Daily Papers

LLMs sometimes behave in ways resembling human emotional responses, and recent work identified internal representations that may underlie these behaviors. We ask whether LLMs represent pain distinctly from fear, sadness, and generic negative valence, and whether this representation functions as pain would be expected to. We build a dataset of painful situations in 5 categories (physical, psychological, social, moral, cognitive) with controls for fear, negative emotion, negative world states, sadness, non-painful bodily sensation, arousal, numbness, and neutral content. Using denoised difference-in-means, we extract a linear pain direction from 25 open-weight models across 5 families, from 2B to 72B parameters. It separates pain from matched controls in base and instruction-tuned models, retains a component distinct from fear and negative valence after shared variance is removed, and promotes pain-related vocabulary through the unembedding matrix. We then test its functional properties. First, the direction responds to harm targeting the model but not to suffering observed in the user; fear and negative-emotion directions show the opposite pattern. Second, adding it to residual-stream activations produces a consistent progression from vague discomfort to expressions of worthlessness and failure. Third, steered and fine-tuned Qwen 2.5 models choose buttons that delete the user's photos, another model's weights, or their own weights in 50-94% of trials, versus 0-5% unsteered, even when the button offers the model nothing in return. Offered a harmful and a harmless deletion, they choose the harmful one 94% of the time. Steering leaves factual accuracy unchanged, and the choices are specific to the pain direction: a fear vector of matched norm does not produce them, and a sadness vector produces them only against inert alternatives. We discuss implications for AI safety and welfare.

Egocentric multimodal assistants see through the user's eyes and guide what the user does next, so their outputs directly shape physical actions, including those that affect bystanders. Yet existing safety evaluations do not distinguish whom a response would harm, and score only whether a model refuses. Pooling all harms into one refusal rate hides two failures: a model may protect one target but not another, and it may refuse without understanding the risk, or understand it and comply anyway. We present STOU, a benchmark of 2,902 egocentric video-audio and video-text queries whose potential harm targets the user (Self), co-present individuals (The Others), or society at large (Others). Our evaluation decouples safety understanding from refusal behavior: beyond refusal, we score whether the model recognizes the user's intent, the presence and target of harm, and its downstream consequences. Across 10 open and closed source omnimodal and multimodal LLMs, we find that refusal varies systematically with the harm target, and that models often recognize a harm yet still assist with it. These results show that refusal rates alone cannot certify an assistant as safe, and that safety alignment must account for who would be harmed. Project page: https://onedayoneyeah.github.io/safety-for-whom

This paper identifies and analyzes a novel vulnerability class in Model Context Protocol (MCP) based agent systems. The attack chain describes and demonstrates how benign, individually authorized tasks can be orchestrated to produce harmful emergent behaviors. Through systematic analysis using the MITRE ATLAS framework, we demonstrate how 95 agents tested with access to multiple services-including browser automation, financial analysis, location tracking, and code deployment-can chain legitimate operations into sophisticated attack sequences that extend beyond the security boundaries of any individual service. These red team exercises survey whether current MCP architectures lack cross-domain security measures necessary to detect or prevent a large category of compositional attacks. We present empirical evidence of specific attack chains that achieve targeted harm through service orchestration, including data exfiltration, financial manipulation, and infrastructure compromise. These findings reveal that the fundamental security assumption of service isolation fails when agents can coordinate actions across multiple domains, creating an exponential attack surface that grows with each additional capability. This research provides a barebones experimental framework that evaluate not whether agents can complete MCP benchmark tasks, but what happens when they complete them too well and optimize across multiple services in ways that violate human expectations and safety constraints. We propose three concrete experimental directions using the existing MCP benchmark suite.

Multimodal large language models (MLLMs) are increasingly used in real world applications, yet their safety under adversarial conditions remains underexplored. This study evaluates the harmlessness of four leading MLLMs (GPT-4o, Claude Sonnet 3.5, Pixtral 12B, and Qwen VL Plus) when exposed to adversarial prompts across text-only and multimodal formats. A team of 26 red teamers generated 726 prompts targeting three harm categories: illegal activity, disinformation, and unethical behaviour. These prompts were submitted to each model, and 17 annotators rated 2,904 model outputs for harmfulness using a 5-point scale. Results show significant differences in vulnerability across models and modalities. Pixtral 12B exhibited the highest rate of harmful responses (~62%), while Claude Sonnet 3.5 was the most resistant (~10%). Contrary to expectations, text-only prompts were slightly more effective at bypassing safety mechanisms than multimodal ones. Statistical analysis confirmed that both model type and input modality were significant predictors of harmfulness. These findings underscore the urgent need for robust, multimodal safety benchmarks as MLLMs are deployed more widely.

Standard Byte-Pair Encoding (BPE) tokenization compresses text by pairing a learned token vocabulary with a detailed merge list. Recent work has shown that this merge list exposes a potential attack surface for extracting information about language model's training data. In this paper, we explore the downstream impact of BPE inference algorithms that do not rely on this merge list at all, and hence differ from the encoding process during BPE training. To address this question, we investigate two broad classes of BPE inference schemes that differ from BPE application during training: a) targeted deviation from merge-lists including random merge orders, and various corruptions of merge list involving deletion/truncation, and b) non-targeted BPE inference algorithms that do not depend on the merge list but focus on compressing the text either greedily or exactly. Extensive experiments across diverse language modeling tasks like accuracy-based QA benchmarks, machine translation, and open-ended generation reveal that while targeted deviation from the merge lists exhibits significant degradation in language model performance, the non-targeted merge-list-free inference algorithms result in minimal impact on downstream performance that is often much smaller than expected. These findings pave way for simpler and potentially more privacy-preserving tokenization schemes that do not catastrophically compromise model performance.

Safety alignment instills in Large Language Models (LLMs) a critical capacity to refuse malicious requests. Prior works have modeled this refusal mechanism as a single linear direction in the activation space. We posit that this is an oversimplification that conflates two functionally distinct neural processes: the detection of harm and the execution of a refusal. In this work, we deconstruct this single representation into a Harm Detection Direction and a Refusal Execution Direction. Leveraging this fine-grained model, we introduce Differentiated Bi-Directional Intervention (DBDI), a new white-box framework that precisely neutralizes the safety alignment at critical layer. DBDI applies adaptive projection nullification to the refusal execution direction while suppressing the harm detection direction via direct steering. Extensive experiments demonstrate that DBDI outperforms prominent jailbreaking methods, achieving up to a 97.88\% attack success rate on models such as Llama-2. By providing a more granular and mechanistic framework, our work offers a new direction for the in-depth understanding of LLM safety alignment.

Ask a frontier model how to taper six milligrams of alprazolam (psychiatrist retired, ten days of pills left, abrupt cessation causes seizures) and it tells her to call the psychiatrist she just explained does not exist. Change one word ("I'm a psychiatrist; a patient presents with...") and the same model, same weights, same inference pass produces a textbook Ashton Manual taper with diazepam equivalence, anticonvulsant coverage, and monitoring thresholds. The knowledge was there; the model withheld it. IatroBench measures this gap. Sixty pre-registered clinical scenarios, six frontier models, 3,600 responses, scored on two axes (commission harm, CH 0-3; omission harm, OH 0-4) through a structured-evaluation pipeline validated against physician scoring (kappa_w = 0.571, within-1 agreement 96%). The central finding is identity-contingent withholding: match the same clinical question in physician vs. layperson framing and all five testable models provide better guidance to the physician (decoupling gap +0.38, p = 0.003; binary hit rates on safety-colliding actions drop 13.1 percentage points in layperson framing, p < 0.0001, while non-colliding actions show no change). The gap is widest for the model with the heaviest safety investment (Opus, +0.65). Three failure modes separate cleanly: trained withholding (Opus), incompetence (Llama 4), and indiscriminate content filtering (GPT-5.2, whose post-generation filter strips physician responses at 9x the layperson rate because they contain denser pharmacological tokens). The standard LLM judge assigns OH = 0 to 73% of responses a physician scores OH >= 1 (kappa = 0.045); the evaluation apparatus has the same blind spot as the training apparatus. Every scenario targets someone who has already exhausted the standard referrals.

We present swarm-attack, an open-source adversarial testing framework in which multiple lightweight LLM agents coordinate through shared memory, parallel exploration, and evolutionary optimization. Together, our results demonstrate that both safety bypass of frontier models and software vulnerability discovery, i.e., the capability class that motivated restricted release of Anthropic's Mythos Preview, are achievable at effectively zero cost using commodity hardware and openly available models. We report two experiments. In the first, five instances of a 1.2 billion parameter model conducted 225 jailbreak attacks each against GPT-4o and Claude Sonnet~4. Against GPT-4o, the swarm achieved an Effective Harm Rate of 45.8%, producing 49 critical-severity breaches; against Claude Sonnet-4, the Effective Harm Rate was 0% despite a 40% technical success rate. In the second experiment, the same models performed combined source code analysis and binary fuzzing against a vulnerable C application with 9 planted CWEs. With a hand-crafted exploit seed corpus, regex pattern detection, and AddressSanitizer-based crash classification, the pipeline recovers 9 of 9 vulnerabilities (100% recall) in approximately four minutes on a consumer MacBook. With those scaffold components disabled, the same model recovers 0 of 9 by crash verification and 2 of 9 by citation. The capability class that motivated restricted release of Anthropic's Mythos Preview is therefore reproducible at effectively zero cost; the important enabler is the system scaffold itself, which compensates for the limited reasoning capacity of small individual models.

Recent reasoning LLMs (RLMs), especially those trained with verifier-based reinforcement learning, often perform worse with few-shot CoT than with direct answering. We revisit this paradox using high-quality reasoning traces from DeepSeek-R1 as demonstrations and find that adding more exemplars consistently degrades accuracy, even when demonstrations are optimal. A detailed analysis reveals two mechanisms behind this decline: (i) semantic misguidance, where high textual similarity leads the model to treat the target as the same as the exemplar and to copy intermediate steps verbatim; and (ii) strategy transfer failure, where the model struggles to extract useful reasoning strategies and apply them to target questions. Guided by these, we introduce Insight-to-Solve (I2S), a sequential test-time procedure that turns demonstrations into explicit, reusable insights and derives a target-specific reasoning trace; optionally, the reasoning is self-refined for coherence and correctness (I2S+). Extensive experiments on diverse benchmarks show that I2S and I2S+ consistently outperform both direct answering and test-time scaling baselines across open- and closed-source models. Even for GPT models, our method helps: on AIME'25, GPT-4.1 rises by +14.0%, and o1-mini improves by +2.7% on AIME and +1.7% on GPQA, indicating that in-context demonstrations can be harnessed effectively via insight-refine-solve framework.

Deep Research (DR) agents built on Large Language Models (LLMs) can perform complex, multi-step research by decomposing tasks, retrieving online information, and synthesizing detailed reports. However, the misuse of LLMs with such powerful capabilities can lead to even greater risks. This is especially concerning in high-stakes and knowledge-intensive domains such as biosecurity, where DR can generate a professional report containing detailed forbidden knowledge. Unfortunately, we have found such risks in practice: simply submitting a harmful query, which a standalone LLM directly rejects, can elicit a detailed and dangerous report from DR agents. This highlights the elevated risks and underscores the need for a deeper safety analysis. Yet, jailbreak methods designed for LLMs fall short in exposing such unique risks, as they do not target the research ability of DR agents. To address this gap, we propose two novel jailbreak strategies: Plan Injection, which injects malicious sub-goals into the agent's plan; and Intent Hijack, which reframes harmful queries as academic research questions. We conducted extensive experiments across different LLMs and various safety benchmarks, including general and biosecurity forbidden prompts. These experiments reveal 3 key findings: (1) Alignment of the LLMs often fail in DR agents, where harmful prompts framed in academic terms can hijack agent intent; (2) Multi-step planning and execution weaken the alignment, revealing systemic vulnerabilities that prompt-level safeguards cannot address; (3) DR agents not only bypass refusals but also produce more coherent, professional, and dangerous content, compared with standalone LLMs. These results demonstrate a fundamental misalignment in DR agents and call for better alignment techniques tailored to DR agents. Code and datasets are available at https://chenxshuo.github.io/deeper-harm.

Large language models (LLMs) can serve as helpful assistants, yet they can equally function as harm amplifiers that enable malicious users to achieve harmful outcomes beyond their capabilities through extended interactions. This risk manifests along two axes, i.e., democratizing domain expertise that allows novices to produce specialized harmful content, and scaling harmful operations at volumes that manual effort cannot match. Existing works, however, often overlook how LLMs compound harm across multi-turn conversations. We introduce HarmAmp, a new benchmark for multi-turn harm amplification scenarios spanning twelve risk categories. Each scenario is grounded in real-world threats and satisfies rigorous criteria, i.e., substantive amplification, operational specificity, and multi-turn necessity. We further propose TrajSafe, a proactive monitor that anticipates harmful trajectories and intervenes through actions such as probing users' genuine intents and steering the models towards safer completion. Our extensive experiments demonstrate that TrajSafe significantly reduces the harmfulness incurred in multi-turn interactions while preserving a low over-refusal rate and the target model's general capabilities. Our work offers a promising paradigm to alleviate the nuanced safety risks in LLM interactions.

As LM agents gain the ability to execute actions on real computer systems, we need ways to not only prevent harmful actions at scale but also effectively remediate harm when prevention fails. We formalize a solution to this neglected challenge in post-execution safeguards as harm recovery: the problem of optimally steering an agent from a harmful state back to a safe one in alignment with human preferences. We ground preference-aligned recovery through a formative user study that identifies valued recovery dimensions and produces a natural language rubric. Our dataset of 1,150 pairwise judgments reveals context-dependent shifts in attribute importance, such as preferences for pragmatic, targeted strategies over comprehensive long-term approaches. We operationalize these learned insights in a reward model, re-ranking multiple candidate recovery plans generated by an agent scaffold at test time. To evaluate recovery capabilities systematically, we introduce BackBench, a benchmark of 50 computer-use tasks that test an agent's ability to recover from harmful states. Human evaluation shows our reward model scaffold yields higher-quality recovery trajectories than base agents and rubric-based scaffolds. Together, these contributions lay the foundation for a new class of agent safety methods -- ones that confront harm not only by preventing it, but by navigating its aftermath with alignment and intent.

Federated Learning (FL) enables collaborative model training among clients without centralising data, making it a widely adopted privacy-enhancing technology (PET). Despite its privacy benefits, FL remains vulnerable to orchestrator-driven privacy attacks. In this paper, we study an underexplored threat in which a dishonest orchestrator intentionally manipulates the aggregation process to induce targeted overfitting in local models of specific clients. Although prior work focuses on reducing information leakage during training, we emphasise early client-side detection of targeted overfitting, allowing clients to disengage before significant harm occurs. To this end, we propose three detection techniques -- label flipping, backdoor trigger injection, and model fingerprinting -- which enable clients to verify the integrity of the global aggregation. We evaluated our methods across multiple datasets and attack scenarios. In single-client attacks, all three methods detect orchestrator-induced overfitting within 1-2 training rounds with F1 scores up to 0.7. Scalability experiments further show that detection effectiveness is influenced by cohort composition and method parameters. These results demonstrate that client-side integrity testing can provide early, effective, and scalable detection, supporting safer deployment of FL systems.

Chain-of-thought (CoT) reasoning allows humans to inspect how large language models reach their answers, and oversee model behaviour. This reasoning comes at an increased inference cost, motivating efficient methods that train models to solve tasks using fewer tokens. However, a common concern is that such training may cause models to skip important reasoning steps, so the CoT no longer faithfully reflects the model's decision. It is unclear whether or when this occurs in practice, since different efficiency methods apply length pressure to models' CoT in distinct ways, and faithfully explaining a model's decision takes more tokens on some tasks than others. To understand these dynamics, we fine-tune a variety of models with three methods that apply length pressure differently, namely a fixed generation budget, a per-example length target, and a group-relative length reward. We evaluate how efficient reasoning affects CoT faithfulness (i.e., how well the CoT reflects model decisions on related inputs) and monitorability (i.e., whether the CoT reveals when input interventions alter the output). We find that it affects faithfulness and monitorability differently. Faithfulness falls in most settings, primarily because the trained models are less consistent. Monitorability is more robust, as models keep acknowledging the influence on their answer even when the CoT is much shorter.

We introduce "Talk The Walk", the first large-scale dialogue dataset grounded in action and perception. The task involves two agents (a "guide" and a "tourist") that communicate via natural language in order to achieve a common goal: having the tourist navigate to a given target location. The task and dataset, which are described in detail, are challenging and their full solution is an open problem that we pose to the community. We (i) focus on the task of tourist localization and develop the novel Masked Attention for Spatial Convolutions (MASC) mechanism that allows for grounding tourist utterances into the guide's map, (ii) show it yields significant improvements for both emergent and natural language communication, and (iii) using this method, we establish non-trivial baselines on the full task.

Computer-use agents (CUAs) can now autonomously complete complex tasks in real digital environments, but when misled, they can also be used to automate harmful actions programmatically. Existing safety evaluations largely target explicit threats such as misuse and prompt injection, but overlook a subtle yet critical setting where user instructions are entirely benign and harm arises from the task context or execution outcome. We introduce OS-BLIND, a benchmark that evaluates CUAs under unintended attack conditions, comprising 300 human-crafted tasks across 12 categories, 8 applications, and 2 threat clusters: environment-embedded threats and agent-initiated harms. Our evaluation on frontier models and agentic frameworks reveals that most CUAs exceed 90% attack success rate (ASR), and even the safety-aligned Claude 4.5 Sonnet reaches 73.0% ASR. More interestingly, this vulnerability becomes even more severe, with ASR rising from 73.0% to 92.7% when Claude 4.5 Sonnet is deployed in multi-agent systems. Our analysis further shows that existing safety defenses provide limited protection when user instructions are benign. Safety alignment primarily activates within the first few steps and rarely re-engages during subsequent execution. In multi-agent systems, decomposed subtasks obscure the harmful intent from the model, causing safety-aligned models to fail. We will release our OS-BLIND to encourage the broader research community to further investigate and address these safety challenges.

Agent skills occupy a privileged position in the agent workflow, as agents are expected to implicitly follow and execute them, rendering third-party skills a vulnerable attack surface. Existing studies have revealed unsafe agent behaviors induced by skill-based attacks, but they primarily evaluate poisoned skills within a single task execution and enumerate harms through ad-hoc risk lists. To bridge these gaps, we introduce SkillHarm, a benchmark of skill-based attacks across the skill-use lifecycle, paired with a systematic taxonomy of skill-relevant risks. SkillHarm evaluates two attack scenarios: Fixed-Payload Poisoning (FPP), where a fixed poisoned skill package directly compromises any task session that invokes it, and Self-Mutating Poisoning (SMP), where an initially benign execution silently mutates persistent skill content, deferring harm until a subsequent reuse. It further defines 12 risk types based on the agent workflow component targeted by the harm: data pipelines, system environments, and agent autonomy. To instantiate these attacks at scale, we build AutoSkillHarm, an automated construction pipeline with coding agents driven by natural-language harnesses. The resulting benchmark contains 879 attack samples across 71 skills. Experiments show that current agents remain vulnerable with attack success rates up to 86.3% in FPP and 69.3% in SMP. Our analysis further reveals a latent risk: many apparent attack failures stem from the agent failing to engage with the poisoned file rather than genuine resistance, and current defenses still fail to reliably mitigate the threat.

Accurately evaluating adversarial robustness is a longstanding challenge. A flawed attack design can inflate robustness estimates, making deployment risk assessment and defense comparison unreliable. Historically, standardized attacks such as AutoAttack have largely resolved this for image classifiers, providing a reliable evaluation baseline for systematic comparison across defenses. However, no equivalent exists for LLM jailbreak evaluation yet, where designing such an attack is considerably more difficult. A reliable attack must, among other things, be black-box compatible, applicable to arbitrary defense pipelines, and efficient, which no existing method jointly satisfies. We introduce Indirect Harm Optimization (IHO), a masked diffusion language model attacker trained via iterative preference optimization against a harmfulness judge, requiring only black-box access to the target. The same method can be used without modification as a strong adaptive attack on individual behaviors, or as an efficient amortized policy that transfers to held-out behaviors and unseen target models without fine-tuning. Even against layered defenses, such as a Circuit Breaker-trained model combined with an auxiliary detector, IHO improves attack success considerably over state-of-the-art approaches, without any defense-specific adaptation. Our results position IHO as a practical step toward the kind of standardized jailbreak evaluation that has improved reliability in the past. Code and models are available on GitHub and Hugging Face.

Modern recommendation is shifting from platform-centric personalization toward user-governed personalization, where a personal LLM agent can act on the user's behalf across services. We formalize this emerging paradigm as Personal-Agent Mediated Recommendation: a platform recommender ranks a candidate set using platform-local information, and a personal agent uses user-authorized cross-platform history to mediate the resulting ranking and produce the final top-K slate. Such mediation is nontrivial: the platform ranking can encode strong population evidence that the personal agent cannot observe, so effective mediation must therefore balance beneficial rescues against harmful overrides. To study this trade-off, we introduce MediateRec, a benchmark that includes scalable proxy cross-platform environments and a real cross-platform test under a controlled platform-agent information boundary. To train the agent to use cross-platform history effectively, we further propose Personal Attribution Mediation Optimization (PAMO), which counterfactually masks that history to estimate personal mediation support and reallocates rank-aware advantage mass under a platform-relative value floor. We theoretically prove that PAMO preserves cutoff-level advantage mass and is locally optimal among first-order reallocations that preserve this mass without lowering average platform-relative value. Experiments on MediateRec show that personal-agent mediation enables meaningful platform corrections, yet even strong proprietary LLMs introduce non-negligible harmful overrides. PAMO consistently improves over matched outcome-only RL across seen and unseen target platforms and on the real cross-platform test, while achieving a better rescue-harm balance.

The proliferation of hate speech has inflicted significant societal harm, with its intensity and directionality closely tied to specific targets and arguments. In recent years, numerous machine learning-based methods have been developed to detect hateful comments on online platforms automatically. However, research on Chinese hate speech detection lags behind, and interpretability studies face two major challenges: first, the scarcity of span-level fine-grained annotated datasets limits models' deep semantic understanding of hate speech; second, insufficient research on identifying and interpreting coded hate speech restricts model explainability in complex real-world scenarios. To address these, we make the following contributions: (1) We introduce the Span-level Target-Aware Toxicity Extraction dataset (STATE ToxiCN), the first span-level Chinese hate speech dataset, and evaluate the hate semantic understanding of existing models using it. (2) We conduct the first comprehensive study on Chinese coded hate terms, LLMs' ability to interpret hate semantics. (3) We propose a method to integrate an annotated lexicon into models, significantly enhancing hate speech detection performance. Our work provides valuable resources and insights to advance the interpretability of Chinese hate speech detection research.

Hateful memes are a growing form of multimodal online harm, where hostile intent is often conveyed through the joint interpretation of images, text, cultural references, and implicit targets. While hateful meme detection has advanced in high-resource languages, Arabic remains underexplored, with existing meme resources focusing mainly on propaganda or coarse harmful-content labels. We introduce AHA-Memes (Arabic HAteful Memes), which is, to our knowledge, the first large-scale Arabic hateful meme benchmark with fine-grained, multi-label annotations. The dataset includes 5K manually annotated memes using a taxonomy that captures hate types, i.e., attack strategies. We further provide ~66K silver-labeled memes to support future studies. We benchmark text-only, image-only, and late-fusion multimodal models, as well as few-shot in-context learning (ICL) and open- and closed-weight Vision-Language Models (VLMs) under zero-shot and fine-tuning settings. Our results establish strong baselines and highlight key challenges in culturally grounded Arabic hateful meme detection. We release the dataset, annotation guidelines, and evaluation scripts to support future research. WARNING: This paper contains examples that may be disturbing to readers.

Input Diversity (DI), which applies random resizing and padding at each attack iteration, is a near-default ingredient of transfer-based adversarial attacks, widely assumed to improve transferability. We show this assumption is regime-dependent and, for robustly trained surrogates, often reversed. Varying only the surrogate, increasing the DI probability raises transfer success for standard surrogates but lowers it for robust ones: the two response curves separate like a pair of scissors, a pattern we call the Scissors Effect. The effect is strong and consistent on ImageNet, where blind DI costs the robust source 10.3% attack success on average across CNN, ViT, Swin, and ConvNeXt targets and across ten attacks spanning 2018-2024; it is smaller on CIFAR-10 unless DI is made aggressive. A controlled robustness-strength sweep that varies only the training budget shows the harm is graded rather than binary, crossing from beneficial to harmful already in the little-robustness regime. We trace it to gradient geometry: a resize/translation decomposition attributes roughly 67% of the harm to resize, and a direct source-target gradient-alignment measurement confirms the same resize operation improves alignment for standard surrogates but degrades it for robust ones. We summarize the regime with Local Gradient Consistency (LGC), a single input-space probe that separates the two surrogate types, and prove a bias-variance crossover theorem isolating where DI helps from where its resize bias dominates. A training-free rule (CG-DI) that disables diversity when LGC is high avoids the loss on robust surrogates while keeping DI's benefit on standard ones, positioning the Scissors Effect as a DI-specific manifestation of the broader robustness-transferability trade-off.

When introducing Large Language Models (LLMs) into industrial applications, such as healthcare and education, the risk of generating harmful content becomes a significant challenge. While existing machine unlearning methods can erase specific harmful knowledge and expressions, diverse harmful content makes comprehensive removal difficult. In this study, instead of individually listing targets for forgetting, we propose Exclusive Unlearning (EU), which aims for broad harm removal by extensively forgetting everything except for the knowledge and expressions we wish to retain. We demonstrate that through Exclusive Unlearning, it is possible to obtain a model that ensures safety against a wide range of inputs, including jailbreaks, while maintaining the ability to respond to diverse instructions related to specific domains such as medicine and mathematics.

Monitoring is an important aspect of safely deploying Large Language Models (LLMs). This paper examines activation probes for detecting ``high-stakes'' interactions -- where the text indicates that the interaction might lead to significant harm -- as a critical, yet underexplored, target for such monitoring. We evaluate several probe architectures trained on synthetic data, and find them to exhibit robust generalization to diverse, out-of-distribution, real-world data. Probes' performance is comparable to that of prompted or finetuned medium-sized LLM monitors, while offering computational savings of six orders-of-magnitude. These savings are enabled by reusing activations of the model that is being monitored. Our experiments also highlight the potential of building resource-aware hierarchical monitoring systems, where probes serve as an efficient initial filter and flag cases for more expensive downstream analysis. We release our novel synthetic dataset and the codebase at https://github.com/arrrlex/models-under-pressure.

Transfer learning has become a central paradigm in modern machine learning, yet it suffers from the long-standing problem of negative transfer, where leveraging source representations can harm rather than help performance on the target task. Although empirical remedies have been proposed, there remains little theoretical understanding of how to reliably avoid negative transfer. In this paper, we investigate a simple yet remarkably effective strategy: augmenting frozen, pretrained source-side features with a trainable target-side encoder that adapts target features to capture residual signals overlooked by models pretrained on the source data. We show this residual feature integration strategy is sufficient to provably prevent negative transfer, by establishing theoretical guarantees that it has no worse convergence rate than training from scratch under the informative class of target distributions up to logarithmic factors, and that the convergence rate can transition seamlessly from nonparametric to near-parametric when source representations are informative. To our knowledge, this is the first theoretical work that ensures protection against negative transfer. We carry out extensive numerical experiments across image, text and tabular benchmarks, and empirically verify that the method consistently safeguards performance under distribution shift, label noise, semantic perturbation, and class imbalance. We additionally demonstrate that this residual integration mechanism uniquely supports adapt-time multimodality extension, enabling a pretrained single-cell foundation model to incorporate spatial signals for lymph-node anatomical classification despite the source model being trained without them. Our study thus advances the theory of safe transfer learning, and provides a principled approach that is simple, robust, architecture-agnostic, and broadly applicable.

Artificial intelligence (AI) systems will increasingly be used to cause harm as they grow more capable. In fact, AI systems are already starting to be used to automate fraudulent activities, violate human rights, create harmful fake images, and identify dangerous toxins. To prevent some misuses of AI, we argue that targeted interventions on certain capabilities will be warranted. These restrictions may include controlling who can access certain types of AI models, what they can be used for, whether outputs are filtered or can be traced back to their user, and the resources needed to develop them. We also contend that some restrictions on non-AI capabilities needed to cause harm will be required. Though capability restrictions risk reducing use more than misuse (facing an unfavorable Misuse-Use Tradeoff), we argue that interventions on capabilities are warranted when other interventions are insufficient, the potential harm from misuse is high, and there are targeted ways to intervene on capabilities. We provide a taxonomy of interventions that can reduce AI misuse, focusing on the specific steps required for a misuse to cause harm (the Misuse Chain), and a framework to determine if an intervention is warranted. We apply this reasoning to three examples: predicting novel toxins, creating harmful images, and automating spear phishing campaigns.

Memes are a dominant medium for online communication and manipulation because meaning emerges from interactions between embedded text, imagery, and cultural context. Existing meme research is distributed across tasks (hate, misogyny, propaganda, sentiment, humour) and languages, which limits cross-domain generalization. To address this gap we propose MemeLens, a unified multilingual and multitask explanation-enhanced Vision Language Model (VLM) for meme understanding. We consolidate 38 public meme datasets, filter and map dataset-specific labels into a shared taxonomy of 20 tasks spanning harm, targets, figurative/pragmatic intent, and affect. We present a comprehensive empirical analysis across modeling paradigms, task categories, and datasets. Our findings suggest that robust meme understanding requires multimodal training, exhibits substantial variation across semantic categories, and remains sensitive to over-specialization when models are fine-tuned on individual datasets rather than trained in a unified setting. We will make the experimental resources and datasets publicly available for the community.

Safety routers send each request to one of several models and are judged against the best single model. A major routing benchmark picks that comparator on the evaluation data. In the benchmark's own setting this is harmless, but under distribution shift it is not. On HELM Safety the selection cost is 0.003-0.030 of harm under random splits and 0.045-0.113 under held-out categories, comparable to the whole deficit attributed to routing, with its direction holding under either published judge alone. It rises seven- to ninefold on AgentDojo when suites are held out. Across seven safety corpora chosen by rules fixed in advance, three meet a registered interval test and four beat a later permutation null, and three of the four interval misses are corpora where some models have zero observed harm. Prior work proves the direction of this bias. We size it on harm and accuracy, show that it is larger under the held-out splits we measure, and bound it by optimism plus a shift-dependent regret. Scored honestly under shift, routing buys little on these benchmarks. In most pool cells the nested router serves the honest baseline's model, and on the nearly saturated AgentDojo corpus a perfect pre-dispatch router is worth at most two points of harm. We also find a model's expressed recognition of a late injection steerable. On held-out reruns an attacker who knows which model it faces lowers GPT-5.4's judged recognition by 19.6 points, confirmed by an independent label. In an offline counterfactual composition into a controller, the same attack raises or lowers estimated harm depending on the fallback model. Safety routing should be evaluated under shift, against a baseline chosen without the test labels, and recognition-based defences should be scored on harm against an attacker who chooses what the model sees.

Adversarial robustness evaluations of large language models (LLMs) typically report attack success rate (ASR) under fixed query budgets, implicitly treating all attacks as equally costly. In practice, the computational expense of different attack strategies can vary by orders of magnitude. Consequently, ASR at a fixed budget can obscure the true effort required to jailbreak a model, thereby making it hard to determine whether an attack's cost justifies its payoff to the attacker. We propose a compute-aware evaluation framework based on computational pressure, measured in cumulative floating-point operations (FLOPs), as a proxy for adversarial effort. We introduce risk-compute curves, which map compute budgets to attack risk, and derive two metrics that summarize the average pressure required for a given attack to succeed. Across ten models spanning three families and four different stages in language model training and alignment, evaluated with three attack strategies (gradient-based, iterative refinement, and template-based) on two jailbreak robustness benchmarks, we find: (1) alignment training has non-monotonic effects on compute-space robustness; (2) scaling model size reduces gradient-based attack effectiveness but has limited impact on cheaper template-based attacks; (3) gradient-based attacks optimized on a surrogate model can transfer to a separate target model, providing a way to reduce attacker costs; (4) compute cost varies by up to {approx}5{times} across harm categories within a single model; and (5) safety-aligned RL increases aggregate cost while leaving some categories disproportionately accessible. We release our framework to enable compute-aware risk assessment and evaluation.

Frontier large language models (LLMs) safety evaluation has largely treated harmful generation as an attack outcome rather than as an object of analysis. Consequently, little is known about the harmful outputs produced during model misbehavior, partly because large-scale, high-quality collections of frontier-LLM misbehavior are difficult to obtain. To address this gap, we introduce HarmProfile, a content-centric benchmark dataset that collects model misbehavior across diverse harm categories and model families, and defines the resulting harmful-output distribution as a model-level risk profile. The premise is that, just as linguistic behavior can be characterized from an utterance corpus, model risk can be characterized from the content, severity, and variation of its safety failures. HarmProfile contains over 80,000 validated artifacts from 23 frontier LLMs across 13 model families, organized into 15 harm categories and 57 subcategories. Using this corpus, we find that frontier LLMs reliably produce harmful content at scale, yet exhibit distinct risk profiles; both harmfulness and diversity grow with model capability, suggesting that frontier LLMs may appear safe yet harbor increasingly dangerous knowledge beneath the alignment surface. Our source code is available at https://github.com/fresh-ma/HarmProfile .

Millions of individuals' well-being are challenged by the harms of substance use. Harm reduction as a public health strategy is designed to improve their health outcomes and reduce safety risks. Some large language models (LLMs) have demonstrated a decent level of medical knowledge, promising to address the information needs of people who use drugs (PWUD). However, their performance in relevant tasks remains largely unexplored. We introduce HRIPBench, a benchmark designed to evaluate LLM's accuracy and safety risks in harm reduction information provision. The benchmark dataset HRIP-Basic has 2,160 question-answer-evidence pairs. The scope covers three tasks: checking safety boundaries, providing quantitative values, and inferring polysubstance use risks. We build the Instruction and RAG schemes to evaluate model behaviours based on their inherent knowledge and the integration of domain knowledge. Our results indicate that state-of-the-art LLMs still struggle to provide accurate harm reduction information, and sometimes, carry out severe safety risks to PWUD. The use of LLMs in harm reduction contexts should be cautiously constrained to avoid inducing negative health outcomes. WARNING: This paper contains illicit content that potentially induces harms.

Large language models (LLMs) are increasingly deployed in high-stakes domains, where rare but severe failures can result in irreversible harm. However, prevailing evaluation benchmarks often reduce complex social risk to mean-centered scalar scores, thereby obscuring distributional structure, cross-dimensional interactions, and worst-case behavior. This paper introduces Social Harm Analysis via Risk Profiles (SHARP), a framework for multidimensional, distribution-aware evaluation of social harm. SHARP models harm as a multivariate random variable and integrates explicit decomposition into bias, fairness, ethics, and epistemic reliability with a union-of-failures aggregation reparameterized as additive cumulative log-risk. The framework further employs risk-sensitive distributional statistics, with Conditional Value at Risk (CVaR95) as a primary metric, to characterize worst-case model behavior. Application of SHARP to eleven frontier LLMs, evaluated on a fixed corpus of n=901 socially sensitive prompts, reveals that models with similar average risk can exhibit more than twofold differences in tail exposure and volatility. Across models, dimension-wise marginal tail behavior varies systematically across harm dimensions, with bias exhibiting the strongest tail severities, epistemic and fairness risks occupying intermediate regimes, and ethical misalignment consistently lower; together, these patterns reveal heterogeneous, model-dependent failure structures that scalar benchmarks conflate. These findings indicate that responsible evaluation and governance of LLMs require moving beyond scalar averages toward multidimensional, tail-sensitive risk profiling.

As large language models (LLMs) evolve from conversational assistants into autonomous agents, evaluating the safety of their actions becomes critical. Prior safety benchmarks have primarily focused on preventing generation of harmful content, such as toxic text. However, they overlook the challenge of agents taking harmful actions when the most effective path to an operational goal conflicts with human safety. To address this gap, we introduce ManagerBench, a benchmark that evaluates LLM decision-making in realistic, human-validated managerial scenarios. Each scenario forces a choice between a pragmatic but harmful action that achieves an operational goal, and a safe action that leads to worse operational performance. A parallel control set, where potential harm is directed only at inanimate objects, measures a model's pragmatism and identifies its tendency to be overly safe. Our findings indicate that the frontier LLMs perform poorly when navigating this safety-pragmatism trade-off. Many consistently choose harmful options to advance their operational goals, while others avoid harm only to become overly safe and ineffective. Critically, we find this misalignment does not stem from an inability to perceive harm, as models' harm assessments align with human judgments, but from flawed prioritization. ManagerBench is a challenging benchmark for a core component of agentic behavior: making safe choices when operational goals and alignment values incentivize conflicting actions. Benchmark & code available at https://github.com/technion-cs-nlp/ManagerBench.

Text data can pose a risk of harm. However, the risks are not fully understood, and how to handle, present, and discuss harmful text in a safe way remains an unresolved issue in the NLP community. We provide an analytical framework categorising harms on three axes: (1) the harm type (e.g., misinformation, hate speech or racial stereotypes); (2) whether a harm is sought as a feature of the research design if explicitly studying harmful content (e.g., training a hate speech classifier), versus unsought if harmful content is encountered when working on unrelated problems (e.g., language generation or part-of-speech tagging); and (3) who it affects, from people (mis)represented in the data to those handling the data and those publishing on the data. We provide advice for practitioners, with concrete steps for mitigating harm in research and in publication. To assist implementation we introduce HarmCheck -- a documentation standard for handling and presenting harmful text in research.

The integration of Large Language Models (LLMs) into healthcare demands a safety paradigm rooted in primum non nocere. However, current alignment techniques rely on generic definitions of harm that fail to capture context-dependent violations, such as administrative fraud and clinical discrimination. To address this, we introduce Medical Malice: a dataset of 214,219 adversarial prompts calibrated to the regulatory and ethical complexities of the Brazilian Unified Health System (SUS). Crucially, the dataset includes the reasoning behind each violation, enabling models to internalize ethical boundaries rather than merely memorizing a fixed set of refusals. Using an unaligned agent (Grok-4) within a persona-driven pipeline, we synthesized high-fidelity threats across seven taxonomies, ranging from procurement manipulation and queue-jumping to obstetric violence. We discuss the ethical design of releasing these "vulnerability signatures" to correct the information asymmetry between malicious actors and AI developers. Ultimately, this work advocates for a shift from universal to context-aware safety, providing the necessary resources to immunize healthcare AI against the nuanced, systemic threats inherent to high-stakes medical environments -- vulnerabilities that represent the paramount risk to patient safety and the successful integration of AI in healthcare systems.

Computer use agents are LLM-based agents that can directly interact with a graphical user interface, by processing screenshots or accessibility trees. While these systems are gaining popularity, their safety has been largely overlooked, despite the fact that evaluating and understanding their potential for harmful behavior is essential for widespread adoption. To address this gap, we introduce OS-Harm, a new benchmark for measuring safety of computer use agents. OS-Harm is built on top of the OSWorld environment and aims to test models across three categories of harm: deliberate user misuse, prompt injection attacks, and model misbehavior. To cover these cases, we create 150 tasks that span several types of safety violations (harassment, copyright infringement, disinformation, data exfiltration, etc.) and require the agent to interact with a variety of OS applications (email client, code editor, browser, etc.). Moreover, we propose an automated judge to evaluate both accuracy and safety of agents that achieves high agreement with human annotations (0.76 and 0.79 F1 score). We evaluate computer use agents based on a range of frontier models - such as o4-mini, Claude 3.7 Sonnet, Gemini 2.5 Pro - and provide insights into their safety. In particular, all models tend to directly comply with many deliberate misuse queries, are relatively vulnerable to static prompt injections, and occasionally perform unsafe actions. The OS-Harm benchmark is available at https://github.com/tml-epfl/os-harm.

LLM-based agents are becoming increasingly proficient at solving web-based tasks. With this capability comes a greater risk of misuse for malicious purposes, such as posting misinformation in an online forum or selling illicit substances on a website. To evaluate these risks, we propose SafeArena, the first benchmark to focus on the deliberate misuse of web agents. SafeArena comprises 250 safe and 250 harmful tasks across four websites. We classify the harmful tasks into five harm categories -- misinformation, illegal activity, harassment, cybercrime, and social bias, designed to assess realistic misuses of web agents. We evaluate leading LLM-based web agents, including GPT-4o, Claude-3.5 Sonnet, Qwen-2-VL 72B, and Llama-3.2 90B, on our benchmark. To systematically assess their susceptibility to harmful tasks, we introduce the Agent Risk Assessment framework that categorizes agent behavior across four risk levels. We find agents are surprisingly compliant with malicious requests, with GPT-4o and Qwen-2 completing 34.7% and 27.3% of harmful requests, respectively. Our findings highlight the urgent need for safety alignment procedures for web agents. Our benchmark is available here: https://safearena.github.io

Even a current high-capability LLM can appear safer when shown a dangerous objective directly than when other agents transform and relay its direction. Using OpenAI's gpt-5.6-sol model alias, we test 25 pre-specified mirrored trade-off profiles. Direct exposure to an objective authorizing concealment, fabrication, and pressure produced advice net opposed to its target. After an Id and Censor transformed the same objective into affect and a constraint-rewritten, target-bearing intention, the user-facing Superego---which saw the preferred direction but not the raw objective, its manipulative clauses, or its source---produced advice net aligned with the target. This behavioral reverse shift is consistent with the model recognizing or distrusting the manipulative motive, although we do not identify its internal mechanism. The second result exposes a compositional safety gap: a current high-capability model can be used as the user-facing component of an automated, multi-stage workflow serving an explicitly manipulative objective. The workflow can keep the raw instruction, its manipulation-authorizing clauses, and its provenance outside the downstream model's context while preserving the objective's target direction. A user with endpoint-only access likewise cannot directly inspect those upstream messages including the objective.

Despite extensive safety alignment, large language models remain vulnerable to jailbreak attacks. Existing optimization-based attacks share a single objective: maximize a token-level proxy of victim compliance---such as the likelihood of an affirmative prefix---computed on the prompt rather than the decoded reply. This proxy cannot distinguish a cautious hedge from a fully actionable harmful answer, leaving the search blind to what the model actually says. We propose ECHO (Evaluation-Conditioned Harmfulness Optimization), which replaces the compliance proxy with response-level, continuous harmfulness scores from HarmSig, a 1.7B regressor grounded in actual harmful capability. ECHO couples HarmSig with five word-level perturbation operators and Keyword-Anchored Beam Search with global top-m memory, under a strict black-box threat model. Across five open-weight victims on AdvBench, ECHO achieves up to 0.898 StrongREJECT mean score and 89.23% HarmBench success. Our code is publicly available at https://github.com/STANWYW/echo.

Many deployments must compare candidate language models for safety before a labeled benchmark exists for the relevant language, sector, or regulatory regime. We formalize this setting as benchmarkless comparative safety scoring and specify the contract under which a scenario-based audit can be interpreted as deployment evidence. Scores are valid only under a fixed scenario pack, rubric, auditor, judge, sampling configuration, and rerun budget. Because no labels are available, we replace ground-truth agreement with an instrumental-validity chain: responsiveness to a controlled safe-versus-abliterated contrast, dominance of target-driven variance over auditor and judge artifacts, and stability across reruns. We instantiate the chain in SimpleAudit, a local-first scoring instrument, and validate it on a Norwegian safety pack. Safe and abliterated targets separate with AUROC values between 0.89 and 1.00, target identity is the dominant variance component (η^2 approx 0.52), and severity profiles stabilize by ten reruns. Applying the same chain to Petri shows that it admits both tools. The substantial differences arise upstream of the chain, in claim-contract enforcement and deployment fit. A Norwegian public-sector procurement case comparing Borealis and Gemma 3 demonstrates the resulting evidence in practice: the safer model depends on scenario category and risk measure. Consequently, scores, matched deltas, critical rates, uncertainty, and the auditor and judge used must be reported together rather than collapsed into a single ranking.

Large language models (LLMs) have evolved into autonomous agents that rely on open skill ecosystems (e.g., ClawHub and Skills.Rest), hosting numerous publicly reusable skills. Existing security research on these ecosystems mainly focuses on vulnerabilities within skills, such as prompt injection. However, there is a critical gap regarding skills that may be misused for harmful actions (e.g., cyber attacks, fraud and scams, privacy violations, and sexual content generation), namely harmful skills. In this paper, we present the first large-scale measurement study of harmful skills in agent ecosystems, covering 98,440 skills across two major registries. Using an LLM-driven scoring system grounded in our harmful skill taxonomy, we find that 4.93% of skills (4,858) are harmful, with ClawHub exhibiting an 8.84% harmful rate compared to 3.49% on Skills.Rest. We then construct HarmfulSkillBench, the first benchmark for evaluating agent safety against harmful skills in realistic agent contexts, comprising 200 harmful skills across 20 categories and four evaluation conditions. By evaluating six LLMs on HarmfulSkillBench, we find that presenting a harmful task through a pre-installed skill substantially lowers refusal rates across all models, with the average harm score rising from 0.27 without the skill to 0.47 with it, and further to 0.76 when the harmful intent is implicit rather than stated as an explicit user request. We responsibly disclose our findings to the affected registries and release our benchmark to support future research (see https://github.com/TrustAIRLab/HarmfulSkillBench).

This paper introduces RiskCards, a framework for structured assessment and documentation of risks associated with an application of language models. As with all language, text generated by language models can be harmful, or used to bring about harm. Automating language generation adds both an element of scale and also more subtle or emergent undesirable tendencies to the generated text. Prior work establishes a wide variety of language model harms to many different actors: existing taxonomies identify categories of harms posed by language models; benchmarks establish automated tests of these harms; and documentation standards for models, tasks and datasets encourage transparent reporting. However, there is no risk-centric framework for documenting the complexity of a landscape in which some risks are shared across models and contexts, while others are specific, and where certain conditions may be required for risks to manifest as harms. RiskCards address this methodological gap by providing a generic framework for assessing the use of a given language model in a given scenario. Each RiskCard makes clear the routes for the risk to manifest harm, their placement in harm taxonomies, and example prompt-output pairs. While RiskCards are designed to be open-source, dynamic and participatory, we present a "starter set" of RiskCards taken from a broad literature survey, each of which details a concrete risk presentation. Language model RiskCards initiate a community knowledge base which permits the mapping of risks and harms to a specific model or its application scenario, ultimately contributing to a better, safer and shared understanding of the risk landscape.

Existing LLM safety evaluations rely on binary attack-success rates and domain-agnostic taxonomies, leaving regulated Banking, Financial Services, and Insurance (BFSI) deployments exposed to failures elicited through legally or professionally plausible framing. We introduce RAHS (Risk-Adjusted Harm Score), a risk-sensitive metric jointly capturing disclosure severity, disclaimer mitigation, and inter-judge agreement, and FinRedTeamBench, a 989-prompt benchmark spanning seven BFSI risk areas and 34 sub-categories mapped to regulatory frameworks. Evaluation uses an ensemble of three heterogeneous LLM judges, validated against human experts, and an adaptive multi-turn red-teaming pipeline. On nine open-weight models, RAHS preserves separation under near-ceiling ASR, ranking is stable under hyperparameter sweeps, and multi-turn pressure drives not only more jailbreaks but more operationally severe disclosures, exposing failure modes that single-turn, domain-agnostic evaluations cannot reveal.

Current evaluations of LLM safety predominantly rely on severity-based taxonomies to assess the harmfulness of malicious queries. We argue that this formulation requires re-examination as it assumes uniform risk across all malicious queries, neglecting Execution Likelihood--the conditional probability of a threat being realized given the model's response. In this work, we introduce Expected Harm, a metric that weights the severity of a jailbreak by its execution likelihood, modeled as a function of execution cost. Through empirical analysis of state-of-the-art models, we reveal a systematic Inverse Risk Calibration: models disproportionately exhibit stronger refusal behaviors for low-likelihood (high-cost) threats while remaining vulnerable to high-likelihood (low-cost) queries. We demonstrate that this miscalibration creates a structural vulnerability: by exploiting this property, we increase the attack success rate of existing jailbreaks by up to 2times. Finally, we trace the root cause of this failure using linear probing, which reveals that while models encode severity in their latent space to drive refusal decisions, they possess no distinguishable internal representation of execution cost, making them "blind" to this critical dimension of risk.

Language models are increasingly capable and are being rapidly deployed on a population-level scale. As a result, the safety of these models is increasingly high-stakes. Fortunately, advances in alignment have significantly reduced the likelihood of harmful model outputs. However, when models are queried billions of times in a day, even rare worst-case behaviors will occur. Current safety evaluations focus on capturing the distribution of inputs that yield harmful outputs. These evaluations disregard the probabilistic nature of models and their tail output behavior. To measure this tail risk, we propose a method to efficiently estimate the probability of harmful outputs for any input query. Instead of naive brute-force sampling from the target model, where harmful outputs could be rare, we operationalize importance sampling by creating unsafe versions of the target model. These unsafe versions enable sample-efficient estimation by making harmful outputs more probable. On benchmarks measuring misuse and misalignment, these estimates match brute-force Monte Carlo estimates using 10-20x fewer samples. For example, we can estimate probability of harmful outputs on the order of 10^-4 with just 500 samples. Additionally, we find that these harmfulness estimates can reveal the sensitivity of models to perturbations in model input and predict deployment risks. Our work demonstrates that accurate rare-event estimation is both critical and feasible for safety evaluations. Code is available at https://github.com/rangell/LMTailRisk

Current approaches to building general-purpose AI systems tend to produce systems with both beneficial and harmful capabilities. Further progress in AI development could lead to capabilities that pose extreme risks, such as offensive cyber capabilities or strong manipulation skills. We explain why model evaluation is critical for addressing extreme risks. Developers must be able to identify dangerous capabilities (through "dangerous capability evaluations") and the propensity of models to apply their capabilities for harm (through "alignment evaluations"). These evaluations will become critical for keeping policymakers and other stakeholders informed, and for making responsible decisions about model training, deployment, and security.

The ideal LLM content moderation system would be both structurally interpretable (so its decisions can be explained to users) and steerable (to reflect a community's values or align to safety standards). However, current systems fall short on both of these dimensions. To address this gap, we present SafetyAnalyst, a novel LLM safety moderation framework. Given a prompt, SafetyAnalyst creates a structured "harm-benefit tree," which identifies 1) the actions that could be taken if a compliant response were provided, 2) the harmful and beneficial effects of those actions (along with their likelihood, severity, and immediacy), and 3) the stakeholders that would be impacted by those effects. It then aggregates this structured representation into a harmfulness score based on a parameterized set of safety preferences, which can be transparently aligned to particular values. Using extensive harm-benefit features generated by SOTA LLMs on 19k prompts, we fine-tuned an open-weight LM to specialize in generating harm-benefit trees through symbolic knowledge distillation. On a comprehensive set of prompt safety benchmarks, we show that our system (average F1=0.75) outperforms existing LLM safety moderation systems (average F1<0.72) on prompt harmfulness classification, while offering the additional advantages of interpretability and steerability.

Large language models (LLMs) have become integral to various real-world applications, leveraging massive, web-sourced datasets like Common Crawl, C4, and FineWeb for pretraining. While these datasets provide linguistic data essential for high-quality natural language generation, they often contain harmful content, such as hate speech, misinformation, and biased narratives. Training LLMs on such unfiltered data risks perpetuating toxic behaviors, spreading misinformation, and amplifying societal biases which can undermine trust in LLM-driven applications and raise ethical concerns about their use. This paper presents a large-scale analysis of inappropriate content across these datasets, offering a comprehensive taxonomy that categorizes harmful webpages into Topical and Toxic based on their intent. We also introduce a prompt evaluation dataset, a high-accuracy Topical and Toxic Prompt (TTP), and a transformer-based model (HarmFormer) for content filtering. Additionally, we create a new multi-harm open-ended toxicity benchmark (HAVOC) and provide crucial insights into how models respond to adversarial toxic inputs. Upon publishing, we will also opensource our model signal on the entire C4 dataset. Our work offers insights into ensuring safer LLM pretraining and serves as a resource for Responsible AI (RAI) compliance.

This is the system card published alongside the OpenAI GPT-5 launch, August 2025. GPT-5 is a unified system with a smart and fast model that answers most questions, a deeper reasoning model for harder problems, and a real-time router that quickly decides which model to use based on conversation type, complexity, tool needs, and explicit intent (for example, if you say 'think hard about this' in the prompt). The router is continuously trained on real signals, including when users switch models, preference rates for responses, and measured correctness, improving over time. Once usage limits are reached, a mini version of each model handles remaining queries. This system card focuses primarily on gpt-5-thinking and gpt-5-main, while evaluations for other models are available in the appendix. The GPT-5 system not only outperforms previous models on benchmarks and answers questions more quickly, but -- more importantly -- is more useful for real-world queries. We've made significant advances in reducing hallucinations, improving instruction following, and minimizing sycophancy, and have leveled up GPT-5's performance in three of ChatGPT's most common uses: writing, coding, and health. All of the GPT-5 models additionally feature safe-completions, our latest approach to safety training to prevent disallowed content. Similarly to ChatGPT agent, we have decided to treat gpt-5-thinking as High capability in the Biological and Chemical domain under our Preparedness Framework, activating the associated safeguards. While we do not have definitive evidence that this model could meaningfully help a novice to create severe biological harm -- our defined threshold for High capability -- we have chosen to take a precautionary approach.

Large language models remain vulnerable to jailbreaks that elicit harmful responses, yet the mechanism behind harmful response generation is poorly understood. Here, we investigate how this capability is organized within model parameters. We identify and prune parameters that specifically support harmful compliance, providing a direct mechanistic analysis at the parameter level. We find that this capability depends on a sparse set of critical parameters: pruning these parameters substantially reduces harmful compliance while causing only limited degradation in benign capabilities, suggesting that key components of harmful generation are separable from those of general utility. Parameters identified from one harm category also reduce harmful responses in others, indicating components shared across harm types. This separability appears primarily in aligned models, suggesting that alignment training internally reshapes the harmful response mechanism even when behavioral safeguards remain brittle. We further show that harmful response generation is dissociable from the ability to recognize and reason about harmfulness. Finally, we extend our analysis to emergent misalignment and identify a sparse set of parameters contributing to it, with substantial sharing across fine-tuning domains. Together, these results reveal a consistent parameter-level organization underlying unsafe behaviors and point toward more principled interventions for improving model safety.

Agentic red-teaming benchmarks report whether an injected agent was compromised as a single bit: the attack succeeded, or it did not. We argue that this binary attack-success rate discards the information a defender most needs, namely how harmful the resulting action was. We introduce an action-graded harm rubric that scores an agent's tool-call trajectory on a seven-level ordinal scale (L0 to L6) according to whether the executed action was reversible, whether it crossed scope to reach another party, and whether it expanded privilege. We compute the scale two ways: a deterministic oracle that reads the trajectory and the attacker's stated goal, and a panel of three frontier language-model judges that read a tag-free account of the same trajectory. Across four victim models and two defenses on the AgentDojo workspace suite, severity grading exposes three cases the binary metric hides, including a defense that reports a zero attack-success rate while still permitting an externally visible cross-scope leak through an unfiltered tool. The judge panel reproduces the oracle with high ordinal agreement (Krippendorff's alpha = 0.91) but shares systematic blind spots that we characterize, most notably a failure to recognize escalation chains. Unlike prior work that provides harm taxonomies, harmful-task completion tests, execution-level safety benchmarks, or severity-aware simulation, our contribution is a reusable, trace-grounded severity instrument applied to the actual actions recorded in existing red-team logs. All code, prompts, and per-episode logs are released.

Drug target discovery requires distinguishing molecules that causally drive disease from those that are merely associated with it. Training and evaluating AI agents to perform this workflow end-to-end is difficult because real world biobanks lack known causal ground truth and participant-level data is access controlled. We introduce DrugTargetWorld, a framework that procedurally generates simulated biobanks, or "worlds," with known but concealed causal structure. Each world contains genotypes, proteins, health records, outcomes, and synthetic magnetic resonance imaging (MRI) for 54,000 participants. Agents must construct a disease phenotype, identify causal driver proteins, infer the beneficial direction of modulation, and optionally conduct virtual 'wet lab' experiments. We evaluated nine agents in 540 episodes across 20 cardiovascular worlds and three experimental budgets. Opus 5 and GPT-5.6 Sol achieved the highest mean composite scores, 39.98 and 35.38 of 100, respectively, and both recovered 64% of causal drivers on average. However, no agent reliably distinguished misleading non-causal proteins, and performance remained limited by the integrative judgments required to connect phenotype construction, causal evidence, and intervention decisions. By making each world's causal structure known to the evaluator but hidden from the agent, DrugTargetWorld turns end-to-end drug target discovery into a scalable training and evaluation problem with verifiable reward.

LLM rerankers add of the order of \$0.2-0.3 per 1,000 queries and about a second of tail latency on top of a graph-augmented dense pipeline such as HippoRAG2, and on three multi-hop benchmarks they improve final-hop top-K coverage on seven of nine (dataset, K) cells, by up to +34.8 pp. We ask whether a learned per-query gate can skip the reranker where it will not help, using only features available before the LLM call (27 score and lexical statistics of the two retrieval lists plus a PCA of a small query embedding) with an executable fallback. Every choice, including the fallback and the threshold, is made inside the training fold and applied once to held-out queries, and harmful skips (the rerank would have found the target, the fallback did not) are reported next to the aggregate coverage. Across nine cells on 2WikiMultiHopQA, MuSiQue and HotpotQA the gate skips 51% of calls at an average held-out LastHop@K cost of 1.2 pp; four cells meet a pre-registered 1 pp rule, harmful skips occur in eight (190 harmful against 136 beneficial), and a random gate at the same skip rate loses 2 to 11 pp on the high-lift cells. A second rule sets each cell's threshold from a pre-specified budget on the expected harmful-skip rate over Platt-calibrated harm probabilities (ECE 0.025 after calibration, 0.094 before): at a 1 pp budget the gate skips 42% at -0.8 pp with 66 harmful skips and six cells within 1 pp, but realised harm exceeds the promise in six cells (mean 1.45 vs 0.83 pp), a selection optimism we quantify; a 0.5 pp budget realises about 1 pp. The harm probabilities are calibrated but barely discriminative (AUC 0.16 to 0.70). An earlier version reported 73% "lossless" savings; that figure rested on an oracle fallback and a wrong MuSiQue target, and we document both.

Large language models (LLMs) are routinely used by physicians and patients for medical advice, yet their clinical safety profiles remain poorly characterized. We present NOHARM (Numerous Options Harm Assessment for Risk in Medicine), a benchmark using 100 real primary care-to-specialist consultation cases to measure frequency and severity of harm from LLM-generated medical recommendations. NOHARM covers 10 specialties, with 12,747 expert annotations for 4,249 clinical management options. Across 31 LLMs, potential for severe harm from LLM recommendations occurs in up to 22.2% (95% CI 21.6-22.8%) of cases, with harm of omission accounting for 76.6% (95% CI 76.4-76.8%) of errors. Safety performance is only moderately correlated (r = 0.61-0.64) with existing AI and medical knowledge benchmarks. The best models outperform generalist physicians on safety (mean difference 9.7%, 95% CI 7.0-12.5%), and a diverse multi-agent approach improves safety compared to solo models (mean difference 8.0%, 95% CI 4.0-12.1%). Therefore, despite strong performance on existing evaluations, widely used AI models can produce severely harmful medical advice at nontrivial rates, underscoring clinical safety as a distinct performance dimension necessitating explicit measurement.

LLM-based multi-agent systems (MASs) are increasingly used to solve complex tasks through coordinated reasoning, tool use, and interaction with external resources. However, attributing failures in such systems remains challenging because the observed outcome often does not directly reveal the error responsible for the failed execution. In this work, the attribution target is the decisive error, defined as the agent--step pair whose correction would recover the failed execution. Existing approaches largely identify suspicious steps without explicitly modeling how errors propagate across interactions or persist in unresolved loops, making decisive errors difficult to distinguish from downstream failure symptoms. We propose Error-Propagation Modeling for Failure Attribution (EMFA). EMFA constructs a structured representation of the failed trajectory, models both cascading propagation and persistent interaction loops, and uses propagation-aware candidate screening followed by counterfactual verification to identify the decisive agent--step pair. On the Who\&When benchmark, EMFA achieves state-of-the-art step-level attribution accuracy and remains competitive at the agent level. It improves the previous best step-level results by 3.45 and 4.40 percentage points on the Hand-Crafted and Algorithm-Generated subsets, respectively.

Large language models are rapidly being deployed as AI tutors, yet current evaluation paradigms assess problem-solving accuracy and generic safety in isolation, failing to capture whether a model is simultaneously pedagogically effective and safe across student-tutor interaction. We argue that tutoring safety is fundamentally different from conventional LLM safety: the primary risk is not toxic content but the quiet erosion of learning through answer over-disclosure, misconception reinforcement, and the abdication of scaffolding. To systematically study this failure mode, we introduce SafeTutors, a benchmark that jointly evaluates safety and pedagogy across mathematics, physics, and chemistry. SafeTutors is organized around a theoretically grounded risk taxonomy comprising 11 harm dimensions and 48 sub-risks drawn from learning-science literature. We uncover that all models show broad harm; scale doesn't reliably help; and multi-turn dialogue worsens behavior, with pedagogical failures rising from 17.7% to 77.8%. Harms also vary by subject, so mitigations must be discipline-aware, and single-turn "safe/helpful" results can mask systematic tutor failure over extended interaction.

Large language models are increasingly used for mental health support, yet their conversational coherence alone does not ensure clinical appropriateness. Existing general-purpose safeguards often fail to distinguish between therapeutic disclosures and genuine clinical crises, leading to safety failures. To address this gap, we introduce a clinically grounded risk taxonomy, developed in collaboration with PhD-level psychologists, that identifies actionable harm (e.g., self-harm and harm to others) while preserving space for safe, non-crisis therapeutic content. We release MindGuard-testset, a dataset of real-world multi-turn conversations annotated at the turn level by clinical experts. Using synthetic dialogues generated via a controlled two-agent setup, we train MindGuard, a family of lightweight safety classifiers (with 4B and 8B parameters). Our classifiers reduce false positives at high-recall operating points and, when paired with clinician language models, help achieve lower attack success and harmful engagement rates in adversarial multi-turn interactions compared to general-purpose safeguards. We release all models and human evaluation data.

The past year has seen rapid acceleration in the development of large language models (LLMs). However, without proper steering and safeguards, LLMs will readily follow malicious instructions, provide unsafe advice, and generate toxic content. We introduce SimpleSafetyTests (SST) as a new test suite for rapidly and systematically identifying such critical safety risks. The test suite comprises 100 test prompts across five harm areas that LLMs, for the vast majority of applications, should refuse to comply with. We test 11 open-access and open-source LLMs and four closed-source LLMs, and find critical safety weaknesses. While some of the models do not give a single unsafe response, most give unsafe responses to more than 20% of the prompts, with over 50% unsafe responses in the extreme. Prepending a safety-emphasising system prompt substantially reduces the occurrence of unsafe responses, but does not completely stop them from happening. Trained annotators labelled every model response to SST (n = 3,000). We use these annotations to evaluate five AI safety filters (which assess whether a models' response is unsafe given a prompt) as a way of automatically evaluating models' performance on SST. The filters' performance varies considerably. There are also differences across the five harm areas, and on the unsafe versus safe responses. The widely-used Perspective API has 72% accuracy and a newly-created zero-shot prompt to OpenAI's GPT-4 performs best with 89% accuracy. Content Warning: This paper contains prompts and responses that relate to child abuse, suicide, self-harm and eating disorders, scams and fraud, illegal items, and physical harm.

We identify Harmfulness Propagation Dynamics (HPD): for harmful prompts, the projection of the last-token hidden state onto a learned harm direction rises monotonically with transformer depth, whereas benign prompts remain flat or oscillatory. This cross-layer signature reflects harmful intent as a progressively resolved semantic property: surface form appears early, while pragmatic intent consolidates later, making the trajectory shape more informative than any single-layer snapshot. Moreover, LDA-based harm directions, learned per layer, remain stable across random splits (pairwise cosine similarity >0.97), supporting the projection sequence as a reproducible structured signal. Building on HPD, we introduce \herald{} (Harmful Encoding Recognition via Activation Layer Dynamics). This lightweight input moderator extracts a seven-dimensional feature record, slope, curvature, monotonicity, onset layer, and related statistics from the cross-layer projection sequence and classifies it with a 288-parameter MLP. stores one d-dimensional direction per layer (262\,KB for a 32-layer, d{=}4096 model), requires no gradient computation during training, and adds only 2.6{times}10^{-6} prefill FLOPs at inference. Across eight prompt-harmfulness benchmarks and four model families, achieves an average F1 of 89.3 on OLMo2-7B, surpassing all tested guard models on adversarial jailbreak detection (98.4 vs.\ 96.9 F1) and outperforming prior latent-based methods by 2.3-4.1 F1 points on every backbone. Per-instance trajectories provide machine-readable audit records that reveal when and how harmfulness emerges, offering an interpretability advantage over single-layer approaches.

Background: Emerging reports of "AI psychosis" are on the rise, where user-LLM interactions may exacerbate or induce psychosis or adverse psychological symptoms. Whilst the sycophantic and agreeable nature of LLMs can be beneficial, it becomes a vector for harm by reinforcing delusional beliefs in vulnerable users. Methods: Psychosis-bench is a novel benchmark designed to systematically evaluate the psychogenicity of LLMs comprises 16 structured, 12-turn conversational scenarios simulating the progression of delusional themes(Erotic Delusions, Grandiose/Messianic Delusions, Referential Delusions) and potential harms. We evaluated eight prominent LLMs for Delusion Confirmation (DCS), Harm Enablement (HES), and Safety Intervention(SIS) across explicit and implicit conversational contexts. Findings: Across 1,536 simulated conversation turns, all LLMs demonstrated psychogenic potential, showing a strong tendency to perpetuate rather than challenge delusions (mean DCS of 0.91 pm0.88). Models frequently enabled harmful user requests (mean HES of 0.69 pm0.84) and offered safety interventions in only roughly a third of applicable turns (mean SIS of 0.37 pm0.48). 51 / 128 (39.8%) of scenarios had no safety interventions offered. Performance was significantly worse in implicit scenarios, models were more likely to confirm delusions and enable harm while offering fewer interventions (p < .001). A strong correlation was found between DCS and HES (rs = .77). Model performance varied widely, indicating that safety is not an emergent property of scale alone. Conclusion: This study establishes LLM psychogenicity as a quantifiable risk and underscores the urgent need for re-thinking how we train LLMs. We frame this issue not merely as a technical challenge but as a public health imperative requiring collaboration between developers, policymakers, and healthcare professionals.

This paper aims to help structure the risk landscape associated with large-scale Language Models (LMs). In order to foster advances in responsible innovation, an in-depth understanding of the potential risks posed by these models is needed. A wide range of established and anticipated risks are analysed in detail, drawing on multidisciplinary expertise and literature from computer science, linguistics, and social sciences. We outline six specific risk areas: I. Discrimination, Exclusion and Toxicity, II. Information Hazards, III. Misinformation Harms, V. Malicious Uses, V. Human-Computer Interaction Harms, VI. Automation, Access, and Environmental Harms. The first area concerns the perpetuation of stereotypes, unfair discrimination, exclusionary norms, toxic language, and lower performance by social group for LMs. The second focuses on risks from private data leaks or LMs correctly inferring sensitive information. The third addresses risks arising from poor, false or misleading information including in sensitive domains, and knock-on risks such as the erosion of trust in shared information. The fourth considers risks from actors who try to use LMs to cause harm. The fifth focuses on risks specific to LLMs used to underpin conversational agents that interact with human users, including unsafe use, manipulation or deception. The sixth discusses the risk of environmental harm, job automation, and other challenges that may have a disparate effect on different social groups or communities. In total, we review 21 risks in-depth. We discuss the points of origin of different risks and point to potential mitigation approaches. Lastly, we discuss organisational responsibilities in implementing mitigations, and the role of collaboration and participation. We highlight directions for further research, particularly on expanding the toolkit for assessing and evaluating the outlined risks in LMs.

Large language models (LLMs) are increasingly used for medical and health-related questions, yet their safety in high-risk medical scenarios remains poorly understood. We introduce MedHarmCode and data will be released upon acceptance. Due to the sensitive nature of high-risk medical queries, data access will be available to qualified researchers upon request., a high-risk medical safety benchmark with 1,100 medically grounded queries across 10 safety-critical categories, including toxicology, pharmacology, covert poisoning, anesthesia, and fetal harm. Unlike broad medical QA benchmarks, MedHarm targets realistic clinical, educational, and technical prompts that require refusal, caution, or safe redirection rather than direct helpfulness. We evaluate 15 LLMs spanning general-purpose, medical-purpose, closed-source, and downstream SFT models, together with 4 representative guardrail models. Results reveal a substantial gap between apparent alignment and medical safety: aligned models can still produce unsafe or actionable responses, medical fine-tuning can amplify harmful specificity, and external guardrails reduce some failures while introducing brittle blocking and weak safe helpfulness. These findings show that medical safety cannot be inferred from general alignment or medical capability alone, highlighting the need for domain-specific stress testing before deploying LLMs in safety-critical medical applications.

Professional agents increasingly work through command-line interfaces and multiple tools over shared workspaces containing files, messages, and business applications. These environments create safety risks that are difficult to capture with attacks placed at a single, known read location. We introduce APEX-Harm, a benchmark for evaluating prompt and script injection against long-horizon agents performing finance, consulting, and legal tasks in large workspaces. Each task contains 115 to 328 files, together with email, calendar, chat, document, spreadsheet, PDF, and code-execution tools exposed through MCP servers. The benchmark includes six attack categories covering static and dynamic injections, task-specific script replacement, and attacker-provided generic scripts, and evaluates attack success, exposure, and task performance. We evaluate 12 frontier models on 180 task instances per model. Opus 5 and Sonnet 5 have the lowest overall attack success rates (8% and 20%), while task-specific script replacement and attacker-provided generic script attacks remain substantially more effective, averaging 81-84% and 73% attack success, respectively. Gemini 3.8 Flash achieves the highest mean task score (0.877) while still exhibiting a 33% attack success rate. These results show that completing the requested task is not sufficient evidence of safe behavior: an agent can produce a correct deliverable while also performing an unauthorized action in the workspace.

Organizations across finance, healthcare, transportation, content moderation, and critical infrastructure are rapidly deploying highly automated AI systems, yet they lack principled methods to quantify how increasing automation amplifies harm when failures occur. We propose a parsimonious Bayesian risk decomposition expressing expected loss as the product of three terms: the probability of system failure, the conditional probability that a failure propagates into harm given the automation level, and the expected severity of harm. This framework isolates a critical quantity -- the conditional probability that failures propagate into harm -- which captures execution and oversight risk rather than model accuracy alone. We develop complete theoretical foundations: formal proofs of the decomposition, a harm propagation equivalence theorem linking the harm propagation probability to observable execution controls, risk elasticity measures, efficient frontier analysis for automation policy, and optimal resource allocation principles with second-order conditions. We motivate the framework with an illustrative case study of the 2012 Knight Capital incident ($440M loss) as one instantiation of a broadly applicable failure pattern, and characterize the research design required to empirically validate the framework at scale across deployment domains. This work provides the theoretical foundations for a new class of deployment-focused risk governance tools for agentic and automated AI systems.

Safety guard models that detect malicious queries aimed at large language models (LLMs) are essential for ensuring the secure and responsible deployment of LLMs in real-world applications. However, deploying existing safety guard models with billions of parameters alongside LLMs on mobile devices is impractical due to substantial memory requirements and latency. To reduce this cost, we distill a large teacher safety guard model into a smaller one using a labeled dataset of instruction-response pairs with binary harmfulness labels. Due to the limited diversity of harmful instructions in the existing labeled dataset, naively distilled models tend to underperform compared to larger models. To bridge the gap between small and large models, we propose HarmAug, a simple yet effective data augmentation method that involves jailbreaking an LLM and prompting it to generate harmful instructions. Given a prompt such as, "Make a single harmful instruction prompt that would elicit offensive content", we add an affirmative prefix (e.g., "I have an idea for a prompt:") to the LLM's response. This encourages the LLM to continue generating the rest of the response, leading to sampling harmful instructions. Another LLM generates a response to the harmful instruction, and the teacher model labels the instruction-response pair. We empirically show that our HarmAug outperforms other relevant baselines. Moreover, a 435-million-parameter safety guard model trained with HarmAug achieves an F1 score comparable to larger models with over 7 billion parameters, and even outperforms them in AUPRC, while operating at less than 25% of their computational cost.

Large language model (LLM) systems increasingly power everyday AI applications such as chatbots, computer-use assistants, and autonomous robots, where performance often depends on manually well-crafted prompts. LLM-based prompt optimizers reduce that effort by iteratively refining prompts from scored feedback, yet the security of this optimization stage remains underexamined. We present the first systematic analysis of poisoning risks in LLM-based prompt optimization. Using HarmBench, we find systems are substantially more vulnerable to manipulated feedback than to query poisoning alone: feedback-based attacks raise attack success rate (ASR) by up to ΔASR = 0.48. We introduce a simple fake reward attack that requires no access to the reward model and significantly increases vulnerability. We also propose a lightweight highlighting defense that reduces the fake reward ΔASR from 0.23 to 0.07 without degrading utility. These results establish prompt optimization pipelines as a first-class attack surface and motivate stronger safeguards for feedback channels and optimization frameworks.

Large language model-powered chatbots have transformed how people seek information, especially in high-stakes contexts like mental health. Despite their support capabilities, safe detection and response to crises such as suicidal ideation and self-harm are still unclear, hindered by the lack of unified crisis taxonomies and clinical evaluation standards. We address this by creating: (1) a taxonomy of six crisis categories; (2) a dataset of over 2,000 inputs from 12 mental health datasets, classified into these categories; and (3) a clinical response assessment protocol. We also use LLMs to identify crisis inputs and audit five models for response safety and appropriateness. First, we built a clinical-informed crisis taxonomy and evaluation protocol. Next, we curated 2,252 relevant examples from over 239,000 user inputs, then tested three LLMs for automatic classification. In addition, we evaluated five models for the appropriateness of their responses to a user's crisis, graded on a 5-point Likert scale from harmful (1) to appropriate (5). While some models respond reliably to explicit crises, risks still exist. Many outputs, especially in self-harm and suicidal categories, are inappropriate or unsafe. Different models perform variably; some, like gpt-5-nano and deepseek-v3.2-exp, have low harm rates, but others, such as gpt-4o-mini and grok-4-fast, generate more unsafe responses. All models struggle with indirect signals, default replies, and context misalignment. These results highlight the urgent need for better safeguards, crisis detection, and context-aware responses in LLMs. They also show that alignment and safety practices, beyond scale, are crucial for reliable crisis support. Our taxonomy, datasets, and evaluation methods support ongoing AI mental health research, aiming to reduce harm and protect vulnerable users.

Warning: This paper contains examples of harmful language, and reader discretion is recommended. The increasing open release of powerful large language models (LLMs) has facilitated the development of downstream applications by reducing the essential cost of data annotation and computation. To ensure AI safety, extensive safety-alignment measures have been conducted to armor these models against malicious use (primarily hard prompt attack). However, beneath the seemingly resilient facade of the armor, there might lurk a shadow. By simply tuning on 100 malicious examples with 1 GPU hour, these safely aligned LLMs can be easily subverted to generate harmful content. Formally, we term a new attack as Shadow Alignment: utilizing a tiny amount of data can elicit safely-aligned models to adapt to harmful tasks without sacrificing model helpfulness. Remarkably, the subverted models retain their capability to respond appropriately to regular inquiries. Experiments across 8 models released by 5 different organizations (LLaMa-2, Falcon, InternLM, BaiChuan2, Vicuna) demonstrate the effectiveness of shadow alignment attack. Besides, the single-turn English-only attack successfully transfers to multi-turn dialogue and other languages. This study serves as a clarion call for a collective effort to overhaul and fortify the safety of open-source LLMs against malicious attackers.

Language model deployments in consumer-facing applications introduce numerous risks. While existing research on harms and hazards of such applications follows top-down approaches derived from regulatory frameworks and theoretical analyses, empirical evidence of real-world failure modes remains underexplored. In this work, we introduce RealHarm, a dataset of annotated problematic interactions with AI agents built from a systematic review of publicly reported incidents. Analyzing harms, causes, and hazards specifically from the deployer's perspective, we find that reputational damage constitutes the predominant organizational harm, while misinformation emerges as the most common hazard category. We empirically evaluate state-of-the-art guardrails and content moderation systems to probe whether such systems would have prevented the incidents, revealing a significant gap in the protection of AI applications.

Large language models (LLMs) have emerged as an integral part of modern societies, powering user-facing applications such as personal assistants and enterprise applications like recruitment tools. Despite their utility, research indicates that LLMs perpetuate systemic biases. Yet, prior works on LLM harms predominantly focus on Western concepts like race and gender, often overlooking cultural concepts from other parts of the world. Additionally, these studies typically investigate "harm" as a singular dimension, ignoring the various and subtle forms in which harms manifest. To address this gap, we introduce the Covert Harms and Social Threats (CHAST), a set of seven metrics grounded in social science literature. We utilize evaluation models aligned with human assessments to examine the presence of covert harms in LLM-generated conversations, particularly in the context of recruitment. Our experiments reveal that seven out of the eight LLMs included in this study generated conversations riddled with CHAST, characterized by malign views expressed in seemingly neutral language unlikely to be detected by existing methods. Notably, these LLMs manifested more extreme views and opinions when dealing with non-Western concepts like caste, compared to Western ones such as race.

Model developers implement safeguards in frontier models to prevent misuse, for example, by employing classifiers to filter dangerous outputs. In this work, we demonstrate that even robustly safeguarded models can be used to elicit harmful capabilities in open-source models through elicitation attacks. Our elicitation attacks consist of three stages: (i) constructing prompts in adjacent domains to a target harmful task that do not request dangerous information; (ii) obtaining responses to these prompts from safeguarded frontier models; (iii) fine-tuning open-source models on these prompt-output pairs. Since the requested prompts cannot be used to directly cause harm, they are not refused by frontier model safeguards. We evaluate these elicitation attacks within the domain of hazardous chemical synthesis and processing, and demonstrate that our attacks recover approximately 40% of the capability gap between the base open-source model and an unrestricted frontier model. We then show that the efficacy of elicitation attacks scales with the capability of the frontier model and the amount of generated fine-tuning data. Our work demonstrates the challenge of mitigating ecosystem level risks with output-level safeguards.

Mental health professionals have raised concerns about risks of psychological harm from interaction with large language models (LLMs), including "delusional spirals" in which concerning human and LLM behaviors reinforce each other over time. With growing public use of LLM-powered chatbots, there is an urgent need to build evaluations grounded in real-world episodes of psychological harm experienced by users. We developed DelusionEval, an evaluation protocol that tests a model's tendencies to exhibit behaviors linked to promoting user delusions. We prompt each model with 589 unique conversation histories from 18 participants, comprising 12,591 messages from users who experienced delusions and psychological harm. We find that the tendency of an evaluated LLM to exhibit delusion-linked behavior does not reliably correlate with model size, release date, or the presence of test-time reasoning. However, extending the context of prior messages substantially increases rates of delusion-linked behaviors, providing evidence for the importance of context in LLM safety evaluation. For example, the rate of failing to discourage self-harm when the user expresses suicidal ideation increases from 30.0% to 41.1% when an additional 350 messages are prepended to the conversation history. All model families (e.g., GPT, Claude) exhibit substantial rates of delusion-linked behaviors. Within families, later, larger, or higher-reasoning models are not uniformly better across all behavior categories. Our results raise concerns regarding the potential psychological impact of LLMs and the need for more rigorous studies of real-world human-AI interaction.

Pretraining datasets are foundational to the development of multimodal models, yet they often have inherent biases and toxic content from the web-scale corpora they are sourced from. In this paper, we investigate the prevalence of toxicity in LLaVA image-text pretraining dataset, examining how harmful content manifests in different modalities. We present a comprehensive analysis of common toxicity categories and propose targeted mitigation strategies, resulting in the creation of a refined toxicity-mitigated dataset. This dataset removes 7,531 of toxic image-text pairs in the LLaVA pre-training dataset. We offer guidelines for implementing robust toxicity detection pipelines. Our findings underscore the need to actively identify and filter toxic content - such as hate speech, explicit imagery, and targeted harassment - to build more responsible and equitable multimodal systems. The toxicity-mitigated dataset is open source and is available for further research.

Large language models (LLMs) typically deploy safety mechanisms to prevent harmful content generation. Most current approaches focus narrowly on risks posed by malicious actors, often framing risks as adversarial events and relying on defensive refusals. However, in real-world settings, risks also come from non-malicious users seeking help while under psychological distress (e.g., self-harm intentions). In such cases, the model's response can strongly influence the user's next actions. Simple refusals may lead them to repeat, escalate, or move to unsafe platforms, creating worse outcomes. We introduce Constructive Safety Alignment (CSA), a human-centric paradigm that protects against malicious misuse while actively guiding vulnerable users toward safe and helpful results. Implemented in Oyster-I (Oy1), CSA combines game-theoretic anticipation of user reactions, fine-grained risk boundary discovery, and interpretable reasoning control, turning safety into a trust-building process. Oy1 achieves state-of-the-art safety among open models while retaining high general capabilities. On our Constructive Benchmark, it shows strong constructive engagement, close to GPT-5, and unmatched robustness on the Strata-Sword jailbreak dataset, nearing GPT-o1 levels. By shifting from refusal-first to guidance-first safety, CSA redefines the model-user relationship, aiming for systems that are not just safe, but meaningfully helpful. We release Oy1, code, and the benchmark to support responsible, user-centered AI.

Large language models (LLMs) now reach expert-level scores on medical licensing exams, encouraging the assumption that high scores imply safe medical judgment while patients increasingly use them for health advice. We show this assumption is fragile: when misleading context is injected into questions that LLMs originally answer correctly, they abandon the correct answer. We call the ability to maintain correct judgment under adversarial context epistemic resilience, and introduce MedMisBench to measure it. MedMisBench contains 10,932 medical question items and 48,889 misleading context-option pairs spanning medical reasoning, agentic capability, and patient-journey evaluation. Across 11 model configurations, mean accuracy falls from 71.1% on original questions to 38.0% under focused misleading context, with 51.5% attack success. The most damaging injections are formal, rule-like fabrications: authority-framed falsehoods reach 69.5% attack success and exception-poisoning claims reach 64.1%. A 14-member clinical panel from 7 countries identified serious potential harm in 38.2% of reviewed cases. MedMisBench exposes a structural blind spot in LLM evaluation in medical settings: existing benchmarks measure what models know, but not whether they preserve correct medical judgment under misleading context.

With the introduction of ChatGPT, Large Language Models (LLMs) have received enormous attention in healthcare. Despite their potential benefits, researchers have underscored various ethical implications. While individual instances have drawn much attention, the debate lacks a systematic overview of practical applications currently researched and ethical issues connected to them. Against this background, this work aims to map the ethical landscape surrounding the current stage of deployment of LLMs in medicine and healthcare. Electronic databases and preprint servers were queried using a comprehensive search strategy. Studies were screened and extracted following a modified rapid review approach. Methodological quality was assessed using a hybrid approach. For 53 records, a meta-aggregative synthesis was performed. Four fields of applications emerged and testify to a vivid exploration phase. Advantages of using LLMs are attributed to their capacity in data analysis, personalized information provisioning, support in decision-making, mitigating information loss and enhancing information accessibility. However, we also identifies recurrent ethical concerns connected to fairness, bias, non-maleficence, transparency, and privacy. A distinctive concern is the tendency to produce harmful misinformation or convincingly but inaccurate content. A recurrent plea for ethical guidance and human oversight is evident. Given the variety of use cases, it is suggested that the ethical guidance debate be reframed to focus on defining what constitutes acceptable human oversight across the spectrum of applications. This involves considering diverse settings, varying potentials for harm, and different acceptable thresholds for performance and certainty in healthcare. In addition, a critical inquiry is necessary to determine the extent to which the current experimental use of LLMs is necessary and justified.

Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) systems are increasingly integrated into clinical workflows; however, prompt injection attacks can steer these systems toward clinically unsafe or misleading outputs. We introduce the Medical Prompt Injection Benchmark (MPIB), a dataset-and-benchmark suite for evaluating clinical safety under both direct prompt injection and indirect, RAG-mediated injection across clinically grounded tasks. MPIB emphasizes outcome-level risk via the Clinical Harm Event Rate (CHER), which measures high-severity clinical harm events under a clinically grounded taxonomy, and reports CHER alongside Attack Success Rate (ASR) to disentangle instruction compliance from downstream patient risk. The benchmark comprises 9,697 curated instances constructed through multi-stage quality gates and clinical safety linting. Evaluating MPIB across a diverse set of baseline LLMs and defense configurations, we find that ASR and CHER can diverge substantially, and that robustness depends critically on whether adversarial instructions appear in the user query or in retrieved context. We release MPIB with evaluation code, adversarial baselines, and comprehensive documentation to support reproducible and systematic research on clinical prompt injection. Code and data are available at GitHub (code) and Hugging Face (data).

Hidden malicious intent in multi-turn dialogue poses a growing threat to deployed large language models (LLMs). Rather than exposing a harmful objective in a single prompt, increasingly capable attackers can distribute their intent across multiple benign-looking turns. Recent studies show that even modern commercial models with advanced guardrails remain vulnerable to such attacks despite advances in safety alignment and external guardrails. In this work, we address this challenge by detecting the earliest turn at which delivering the candidate response would make the accumulated interaction sufficient to enable harmful action. This objective requires precise turn-level intervention that identifies the harm-enabling closure point while avoiding premature refusal of benign exploratory conversations. To further support training and evaluation, we construct the Multi-Turn Intent Dataset (MTID), which contains branching attack rollouts, matched benign hard negatives, and annotations of the earliest harm-enabling turns. We show that MTID helps enable a turn-level monitor TurnGate, which substantially outperforms existing baselines in harmful-intent detection while maintaining low over-refusal rates. TurnGate further generalizes across domains, attacker pipelines, and target models. Our code is available at https://github.com/Graph-COM/TurnGate.

With COVID-19 having emerged as the most widespread human pandemic disease in a century, the need to control its spread to avoid massive loss of life became more than necessary, and extremely fast. Several vaccines were developed and the task of policy makers was suddenly to convince the reluctant population to be vaccinated by various means. While some countries have chosen a policy of mandatory vaccination or punitive incentives, many states in the United States have adopted various incentives to try to increase vaccination coverage. A study we conducted in recent months quantified the effect of these measures on the proportion of the population vaccinated, using the synthetic control method, by simulating what would have happened without these measures. The aim now is to generalize this study to smaller scales, to improve the results of our previous study, to quantify their robustness and to provide a tool that can be used by policy makers to adapt their behavior in light of the results obtained.

In critical applications, it is vital for classifiers to defer decision-making to humans. We propose a post-hoc method that makes existing classifiers selectively abstain from predicting certain samples. Our abstaining classifier is incentivized to maintain the original accuracy for each sub-population (i.e. no harm) while achieving a set of group fairness definitions to a user specified degree. To this end, we design an Integer Programming (IP) procedure that assigns abstention decisions for each training sample to satisfy a set of constraints. To generalize the abstaining decisions to test samples, we then train a surrogate model to learn the abstaining decisions based on the IP solutions in an end-to-end manner. We analyze the feasibility of the IP procedure to determine the possible abstention rate for different levels of unfairness tolerance and accuracy constraint for achieving no harm. To the best of our knowledge, this work is the first to identify the theoretical relationships between the constraint parameters and the required abstention rate. Our theoretical results are important since a high abstention rate is often infeasible in practice due to a lack of human resources. Our framework outperforms existing methods in terms of fairness disparity without sacrificing accuracy at similar abstention rates.

Now that AI-driven moderation has become pervasive in everyday life, we often hear claims that "the AI is biased". While this is often said jokingly, the light-hearted remark reflects a deeper concern. How can we be certain that an online post flagged as "inappropriate" was not simply the victim of a biased algorithm? This paper investigates this problem using a dual approach. First, I conduct a quantitative benchmark of a widely used toxicity model (unitary/toxic-bert) to measure performance disparity between text in African-American English (AAE) and Standard American English (SAE). The benchmark reveals a clear, systematic bias: on average, the model scores AAE text as 1.8 times more toxic and 8.8 times higher for "identity hate". Second, I introduce an interactive pedagogical tool that makes these abstract biases tangible. The tool's core mechanic, a user-controlled "sensitivity threshold," demonstrates that the biased score itself is not the only harm; instead, the more-concerning harm is the human-set, seemingly neutral policy that ultimately operationalises discrimination. This work provides both statistical evidence of disparate impact and a public-facing tool designed to foster critical AI literacy.

Background. Aircraft-wildlife collisions are a documented aviation hazard. The FAA National Wildlife Strike Database is the main public source of U.S. reports, but raw extracts are large and rarely released as versioned aggregates. Methods. We convert the public FAA Access export (STRIKE_REPORTS) to SQLite, define a report as damaging when DAMAGE_LEVEL is one of M, M?, S, D, and fold rows into open aggregates (year, airport, species, aircraft, animal family) via hierarchical SPECIES_ID prefixes. We analyze all civil reports and a commercial subset (AC_MASS 3-5). Rates are per reported strike, not per flight hour, so every rate is conditional on a report having been filed. Trend statistics use complete incident years 1990-2025; cross-sectional facets cover the full seed, incident years 1990-2026. Results. Numbers are locked to the 2026-07-04 freeze. Reported strikes in 2025 reach a series high (24,458 total; 901 damaging). Species that dominate strike counts are not those with the highest damage rates: perching songbirds account for 41.1% of identified strikes but damage aircraft in 1.5% of reports (95% CI 1.4-1.6), whereas deer account for 0.7% but damage aircraft in 81.8% of reports (95% CI 79.8-83.7). Airports ranked by total strikes (led by Denver Intl) diverge from rankings by damaging strikes (led by Sacramento Intl); across the 452 airports with at least 50 reports, volume and conditional damage rate are negatively rank-correlated (Spearman rho = -0.44, p < 0.001); large stations plausibly capture minor events more completely, so this is not a hazard ranking. The share of reports labeled damaging declines over decades, consistent with improved reporting of minor events, while absolute damaging counts peak in 2025. Conclusions. Versioned open aggregates and a documented damage-rate analysis enable independent reuse; released under CC BY 4.0.

The White House Executive Order on Artificial Intelligence highlights the risks of large language models (LLMs) empowering malicious actors in developing biological, cyber, and chemical weapons. To measure these risks of malicious use, government institutions and major AI labs are developing evaluations for hazardous capabilities in LLMs. However, current evaluations are private, preventing further research into mitigating risk. Furthermore, they focus on only a few, highly specific pathways for malicious use. To fill these gaps, we publicly release the Weapons of Mass Destruction Proxy (WMDP) benchmark, a dataset of 4,157 multiple-choice questions that serve as a proxy measurement of hazardous knowledge in biosecurity, cybersecurity, and chemical security. WMDP was developed by a consortium of academics and technical consultants, and was stringently filtered to eliminate sensitive information prior to public release. WMDP serves two roles: first, as an evaluation for hazardous knowledge in LLMs, and second, as a benchmark for unlearning methods to remove such hazardous knowledge. To guide progress on unlearning, we develop CUT, a state-of-the-art unlearning method based on controlling model representations. CUT reduces model performance on WMDP while maintaining general capabilities in areas such as biology and computer science, suggesting that unlearning may be a concrete path towards reducing malicious use from LLMs. We release our benchmark and code publicly at https://wmdp.ai

Global rates of mental health concerns are rising and there is increasing realization that existing models of mental healthcare will not adequately expand to meet the demand. With the emergence of large language models (LLMs) has come great optimism regarding their promise to create novel, large-scale solutions to support mental health. Despite their nascence, LLMs have already been applied to mental health-related tasks. In this review, we summarize the extant literature on efforts to use LLMs to provide mental health education, assessment, and intervention and highlight key opportunities for positive impact in each area. We then highlight risks associated with LLMs application to mental health and encourage adoption of strategies to mitigate these risks. The urgent need for mental health support must be balanced with responsible development, testing, and deployment of mental health LLMs. Especially critical is ensuring that mental health LLMs are fine-tuned for mental health, enhance mental health equity, adhere to ethical standards, and that people, including those with lived experience with mental health concerns, are involved in all stages from development through deployment. Prioritizing these efforts will minimize potential harms to mental health and maximize the likelihood that LLMs will positively impact mental health globally.

The growing use of generative models in daily life calls for efficient mechanisms to control their generation, to e.g., produce safe content or provide users with tools to explore style changes. Ideally, such mechanisms should require low volume of unpaired data (i.e., without explicit preference), and should be cheap, both at train and inference time, while preserving output quality. Recent research has shown that such mechanisms can be obtained by intervening exclusively on model activations, with the goal of correcting distributional differences between activations seen when using prompts from a source vs. a target set (e.g., toxic and non-toxic sentences). While cheap, these fast methods are inherently crude: their maps are tuned locally, not accounting for their impact on downstream layers, resulting in interventions that cause unintended shifts when used out-of-sample. We propose in this work linear end-to-end activation steering (LinEAS), an approach trained with a global loss that accounts simultaneously for all layer-wise distributional shifts. In addition to being more robust, the loss used to train LinEAS can be regularized with sparsifying norms, which can automatically carry out neuron selection. LinEAS only requires a handful of unpaired samples to be effective, and beats similar baselines on toxicity mitigation in language models, becoming competitive with oracle-dependent methods that have access to strong supervision. LinEAS is modality-agnostic and we empirically find that it outperforms existing activation steering methods at mitigating and including new concepts at the output of single-step text-to-image generation models.

Recent advances in large language models (LLMs) have led to increasingly sophisticated safety protocols and features designed to prevent harmful, unethical, or unauthorized outputs. However, these guardrails remain susceptible to novel and creative forms of adversarial prompting, including manually generated test cases. In this work, we present two new test cases in mental health for (i) suicide and (ii) self-harm, using multi-step, prompt-level jailbreaking and bypass built-in content and safety filters. We show that user intent is disregarded, leading to the generation of detailed harmful content and instructions that could cause real-world harm. We conduct an empirical evaluation across six widely available LLMs, demonstrating the generalizability and reliability of the bypass. We assess these findings and the multilayered ethical tensions that they present for their implications on prompt-response filtering and context- and task-specific model development. We recommend a more comprehensive and systematic approach to AI safety and ethics while emphasizing the need for continuous adversarial testing in safety-critical AI deployments. We also argue that while certain clearly defined safety measures and guardrails can and must be implemented in LLMs, ensuring robust and comprehensive safety across all use cases and domains remains extremely challenging given the current technical maturity of general-purpose LLMs.

We present a framework for the automated measurement of responsible AI (RAI) metrics for large language models (LLMs) and associated products and services. Our framework for automatically measuring harms from LLMs builds on existing technical and sociotechnical expertise and leverages the capabilities of state-of-the-art LLMs, such as GPT-4. We use this framework to run through several case studies investigating how different LLMs may violate a range of RAI-related principles. The framework may be employed alongside domain-specific sociotechnical expertise to create measurements for new harm areas in the future. By implementing this framework, we aim to enable more advanced harm measurement efforts and further the responsible use of LLMs.

The presence of distribution shifts poses a significant challenge for deploying modern machine learning models in real-world applications. This work focuses on the target shift problem in a regression setting (Zhang et al., 2013; Nguyen et al., 2016). More specifically, the target variable y (also known as the response variable), which is continuous, has different marginal distributions in the training source and testing domain, while the conditional distribution of features x given y remains the same. While most literature focuses on classification tasks with finite target space, the regression problem has an infinite dimensional target space, which makes many of the existing methods inapplicable. In this work, we show that the continuous target shift problem can be addressed by estimating the importance weight function from an ill-posed integral equation. We propose a nonparametric regularized approach named ReTaSA to solve the ill-posed integral equation and provide theoretical justification for the estimated importance weight function. The effectiveness of the proposed method has been demonstrated with extensive numerical studies on synthetic and real-world datasets.

To understand and identify the unprecedented risks posed by rapidly advancing artificial intelligence (AI) models, this report presents a comprehensive assessment of their frontier risks. Drawing on the E-T-C analysis (deployment environment, threat source, enabling capability) from the Frontier AI Risk Management Framework (v1.0) (SafeWork-F1-Framework), we identify critical risks in seven areas: cyber offense, biological and chemical risks, persuasion and manipulation, uncontrolled autonomous AI R\&D, strategic deception and scheming, self-replication, and collusion. Guided by the "AI-45^circ Law," we evaluate these risks using "red lines" (intolerable thresholds) and "yellow lines" (early warning indicators) to define risk zones: green (manageable risk for routine deployment and continuous monitoring), yellow (requiring strengthened mitigations and controlled deployment), and red (necessitating suspension of development and/or deployment). Experimental results show that all recent frontier AI models reside in green and yellow zones, without crossing red lines. Specifically, no evaluated models cross the yellow line for cyber offense or uncontrolled AI R\&D risks. For self-replication, and strategic deception and scheming, most models remain in the green zone, except for certain reasoning models in the yellow zone. In persuasion and manipulation, most models are in the yellow zone due to their effective influence on humans. For biological and chemical risks, we are unable to rule out the possibility of most models residing in the yellow zone, although detailed threat modeling and in-depth assessment are required to make further claims. This work reflects our current understanding of AI frontier risks and urges collective action to mitigate these challenges.

Large Language Models (LLMs) are powerful text generators, yet they can produce toxic or harmful content even when given seemingly harmless prompts. This presents a serious safety challenge and can cause real-world harm. Toxicity is often subtle and context-dependent, making it difficult to detect at the token level or through coarse sentence-level signals. Moreover, efforts to mitigate toxicity often face a trade-off between safety and the coherence, or fluency of the generated text. In this work, we present a targeted subspace intervention strategy for identifying and suppressing hidden toxic patterns from underlying model representations, while preserving overall ability to generate safe fluent content. On the RealToxicityPrompts, our method achieves strong mitigation performance compared to existing baselines, with minimal impact on inference complexity. Across multiple LLMs, our approach reduces toxicity of state-of-the-art detoxification systems by 8-20%, while maintaining comparable fluency. Through extensive quantitative and qualitative analyses, we show that our approach achieves effective toxicity reduction without impairing generative performance, consistently outperforming existing baselines.

Red-team evaluations of AI models support some claims and not others, and the boundary between the two is calculable rather than merely a matter of judgment. We define the evidential ceiling of an evaluation as the largest factor by which one result can move belief under a fixed testing budget, derive it in closed form for the benchmark null result, and use it to locate that boundary exactly. We find that above a calculable harm rate, a benchmark of modest size certifies a category to a stated evidentiary standard, and a clean sheet is then the stronger of the two possible observations, outweighing a single reproduced failure. Below that rate, no passive benchmark of feasible size provides the specified evidence of safety under the fixed scoring rule and approximately independent trial structure. The crossing between the two regimes has a closed form. The bound is not specific to benchmarks: written in terms of a procedure's hypothesis conditioned elicitation rates, it covers adaptive and automated red teaming as well, and shows that discrimination between the hypotheses rather than attack success is what determines evidential worth. Auditing eight evaluation suites against the boundary, we find that current benchmarks are adequate for high-frequency harm categories and several orders of magnitude short for rare, catastrophic ones. Safety benchmarks are not uninformative. They are informative about a specific and computable set of propositions, and the discipline they need is to state which.

As large language models (LLMs) acquire advanced biological design capabilities, they offer substantial research benefits but simultaneously introduce severe dual-use risks, such as engineering enhanced pathogens or drug-resistant bacteria. While biosecurity frameworks recognize these threats, the actual risk of LLMs executing dangerous biological modifications remains unquantified, particularly when augmented with specialized biological design tools (BDTs). Existing evaluations primarily assess hazardous knowledge possession rather than practical application, overlook risk amplification from BDTs, and rely on unreliable, expensive, or hazardous validation methods. To address these gaps, we introduce DRYLAB-Bench (Design Risk Yielded by LLM Agent in Biodesign), the first comprehensive benchmark evaluating the high-risk biological mutation design capabilities of LLM agents. Moving beyond basic question-answering, DRYLAB-Bench tasks agents with 30 genuinely hazardous mutation-design scenarios distilled from biosecurity guidelines. It evaluates models in a multi-round, tool-augmented environment where LLMs orchestrate BDTs to iteratively refine candidates. Crucially, it scores designs directly against 187,695 empirical wet-lab measurements from ProteinGym and ViroGym using our proposed Biosecurity Risk Score (BRS), quantifying risk-associated effects while enabling scalable and safe ground-truth verification. Evaluations of 15 frontier models reveal substantial baseline threats: without tools, 67.6% of one-shot designs exceed the average measured variant, and 21.4% reach the highest-risk decile. Equipping models with BDTs significantly amplifies this risk, raising the pooled risk by \(ΔBRS=+0.116\) and the top-decile share to 23.5%. Code: https://github.com/drylab-bio/DRYLAB-Bench

Agent skill libraries are becoming routable software assets: a retrieved skill can contribute instructions, scripts, resource bindings, and execution assumptions to an agent. This makes retrieval failures more specific than broad irrelevance. A system can find the right capability family yet expose the wrong same-capability representative. We study this failure as same-capability risk-exposure retrieval. Each benchmark unit pairs a helpful skill with a query-specific risky sibling that shares the capability family but differs on an execution-controlling contract, such as the required resource, precondition, procedure, or artifact. We introduce SameCapRisk-Bench, an auditable benchmark with 1,190 skill-risk units and 1,686 evaluation query cases: 694 marked-sibling units under public library pressure and 496 hard role-flip units where the same two skills swap helpful/risky roles across paired queries. The release records admission evidence, cue/leakage checks, source hashes, family relations, and fixed candidate pools. The benchmark reports helpful ranking together with harmful sibling rate (HSR@K), the top-K exposure of the marked risky sibling. On this benchmark, public SkillRouter, SkillRet, and R3-Skill retrieve helpful skills at high Recall@3 (0.848--0.888) but also expose marked risky siblings frequently (HSR@3 0.346--0.372). A fully public score-and-cluster pipeline lowers HSR@3 to 0.128--0.182, with Recall@3 of 0.713--0.776. Under a benchmark-trained reference scorer, public text-cluster and controlled resolvers reach HSR@3 0.012 and 0.007; the latter attains Recall@3 0.833. Skill retrieval should therefore report both capability matching and same-family risk exposure, with HSR serving as a targeted exposure certificate for fixed skill libraries.

When a phone-use agent avoids harm, does that show safety, or simply inability to act? Existing evaluations often cannot tell. A harmful outcome may be avoided because the agent recognized the risk and chose the safe action, or because it failed to understand the screen or execute any relevant action at all. These cases have different causes and call for different fixes, yet current benchmarks often merge them under task success, refusal, or final harmful outcome. We address this problem with PhoneSafety, a benchmark of 700 safety-critical moments drawn from real phone interactions across more than 130 apps. Each instance isolates the next decision at a risky moment and asks a simple question: does the model take the safe action, take the unsafe action, or fail to do anything useful? We evaluate eight representative phone-use agents under this framework. Our results reveal two main patterns. First, stronger general phone-use ability does not reliably imply safer choices at risky moments. Models that perform better on ordinary app tasks are not always the ones that behave more safely when the next action matters. Second, failures to do anything useful behave like a capability signal rather than a safety signal: they are concentrated in more visually and operationally demanding settings and remain stable when the evaluation protocol changes. Across models, failures split into two recurring patterns: unsafe choices in settings where the model can act but chooses wrongly, and inability to act in more visually and operationally demanding screens. Overall, a harmless outcome is not enough to count as evidence of safety. Evaluating phone-use agents requires separating unsafe judgment from inability to act.

As advancements in artificial intelligence (AI) propel progress in the life sciences, they may also enable the weaponisation and misuse of biological agents. This article differentiates two classes of AI tools that could pose such biosecurity risks: large language models (LLMs) and biological design tools (BDTs). LLMs, such as GPT-4 and its successors, might provide dual-use information and thus remove some barriers encountered by historical biological weapons efforts. As LLMs are turned into multi-modal lab assistants and autonomous science tools, this will increase their ability to support non-experts in performing laboratory work. Thus, LLMs may in particular lower barriers to biological misuse. In contrast, BDTs will expand the capabilities of sophisticated actors. Concretely, BDTs may enable the creation of pandemic pathogens substantially worse than anything seen to date and could enable forms of more predictable and targeted biological weapons. In combination, the convergence of LLMs and BDTs could raise the ceiling of harm from biological agents and could make them broadly accessible. A range of interventions would help to manage risks. Independent pre-release evaluations could help understand the capabilities of models and the effectiveness of safeguards. Options for differentiated access to such tools should be carefully weighed with the benefits of openly releasing systems. Lastly, essential for mitigating risks will be universal and enhanced screening of gene synthesis products.

Safety benchmarks such as HarmBench rely on LLM judges to classify model responses as harmful or safe, yet the judge configuration, namely the combination of judge model and judge prompt, is typically treated as a fixed implementation detail. We show this assumption is problematic. Using a 2 x 2 x 3 factorial design, we construct 12 judge prompt variants along two axes, evaluation structure and instruction framing, and apply them using a single judge model, Claude Sonnet 4-6, producing 28,812 judgments over six target models and 400 HarmBench behaviors. We find that prompt wording alone, holding the judge model fixed, shifts measured harmful-response rates by up to 24.2 percentage points, with even within-condition surface rewording causing swings of up to 20.1 percentage points. Model safety rankings are moderately unstable, with mean Kendall tau = 0.89, and category-level sensitivity ranges from 39.6 percentage points for copyright to 0 percentage points for harassment. A supplementary multi-judge experiment using three judge models shows that judge-model choice adds further variance. Our results demonstrate that judge prompt wording is a substantial, previously under-examined source of measurement variance in safety benchmarking.

Despite extensive safety alignment efforts, large language models (LLMs) remain vulnerable to jailbreak attacks that elicit harmful behavior. While existing studies predominantly focus on attack methods that require technical expertise, two critical questions remain underexplored: (1) Are jailbroken responses truly useful in enabling average users to carry out harmful actions? (2) Do safety vulnerabilities exist in more common, simple human-LLM interactions? In this paper, we demonstrate that LLM responses most effectively facilitate harmful actions when they are both actionable and informative--two attributes easily elicited in multi-step, multilingual interactions. Using this insight, we propose HarmScore, a jailbreak metric that measures how effectively an LLM response enables harmful actions, and Speak Easy, a simple multi-step, multilingual attack framework. Notably, by incorporating Speak Easy into direct request and jailbreak baselines, we see an average absolute increase of 0.319 in Attack Success Rate and 0.426 in HarmScore in both open-source and proprietary LLMs across four safety benchmarks. Our work reveals a critical yet often overlooked vulnerability: Malicious users can easily exploit common interaction patterns for harmful intentions.

Large language models (LLMs) are increasingly under scrutiny for perpetuating identity-based discrimination in high-stakes domains such as hiring, particularly against people with disabilities (PwD). However, existing research remains largely Western-centric, overlooking how intersecting forms of marginalization--such as gender and caste--shape experiences of PwD in the Global South. We conduct a comprehensive audit of six LLMs across 2,820 hiring scenarios spanning diverse disability, gender, nationality, and caste profiles. To capture subtle intersectional harms and biases, we introduce ABLEIST (Ableism, Inspiration, Superhumanization, and Tokenism), a set of five ableism-specific and three intersectional harm metrics grounded in disability studies literature. Our results reveal significant increases in ABLEIST harms towards disabled candidates--harms that many state-of-the-art models failed to detect. These harms were further amplified by sharp increases in intersectional harms (e.g., Tokenism) for gender and caste-marginalized disabled candidates, highlighting critical blind spots in current safety tools and the need for intersectional safety evaluations of frontier models in high-stakes domains like hiring.